An industrial pump fault feature extraction method based on time series data analysis

Abstract

The application discloses a kind of industrial pump fault feature extraction method based on time series data analysis, belong to industrial pump fault diagnosis technical field.This method obtains the vibration acceleration signal and outlet pressure signal of multiple monitoring points in industrial pump operation period composition original time series data sequence;Based on the vibration acceleration value and outlet pressure value of each sampling point, construct two-dimensional state vector and arrange as state transition matrix according to time;State transition matrix is singular value decomposed and main singular value sequence is extracted;Difference operation is carried out to main singular value sequence, and the time when the sign of difference singular value changes is marked as candidate mutation time;Local state sub-matrix in the preset time window is intercepted with candidate mutation time as center, the rank of local state sub-matrix is calculated, and the candidate mutation time whose rank is greater than preset threshold is determined as fault feature time;The two-dimensional state vector corresponding to fault feature time is used as fault feature vector, and fault feature set is output.

Description

Technical Field

[0001] This invention relates to the field of industrial pump fault diagnosis technology, specifically to a method for extracting industrial pump fault features based on time-series data analysis. Background Technology

[0002] As core power equipment in process industries, the operating status of industrial pumps directly affects the safety and efficiency of the entire production system. During the long-term operation of industrial pumps, faults such as bearing wear, impeller damage, and seal failure leave characteristic traces in time-series signals in the form of vibration and pressure fluctuations. Accurately extracting characteristic information reflecting the timing of fault occurrence from monitored time-series data is the foundation for early fault warning and diagnosis of industrial pumps.

[0003] Existing methods for extracting fault features from industrial pumps mostly rely on time-domain statistics or spectral analysis of a single physical quantity, such as setting threshold values ​​for vibration signals or analyzing changes in specific frequency components. These methods neglect the coupling relationship between vibration and multidimensional parameters like outlet pressure. During the fault induction phase of industrial pumps, multidimensional parameters often undergo coordinated abrupt changes, and local anomalies in a single channel are easily masked by operating condition fluctuations or background noise, resulting in low temporal accuracy of extracted fault features and even missed detections.

[0004] While anomaly detection methods utilizing multivariate state-space modeling introduce multidimensional information, conventional approaches involving phase space reconstruction or predictive model building suffer from computational complexity and model parameter sensitivity. In locating abrupt changes, simple sliding window statistical differences can easily misidentify sporadic, non-fault-related disturbances as abrupt changes, generating numerous spurious moments and reducing the effectiveness of subsequent fault analysis. When random impulses or measurement interference exist in the monitoring signal, directly searching for abrupt changes from abnormal jumps in the difference or residual sequences makes it difficult to distinguish between changes in dynamic characteristics caused by genuine faults and spurious changes caused by transient disturbances.

[0005] The current problem to be solved is how to obtain the ability to detect mutations that are sensitive to changes in system state and can resist the influence of random disturbances from time series analysis based on the dynamic correlation characteristics of multidimensional monitoring data, and to exclude false mutations from candidate moments and accurately pinpoint the characteristic moments and state vectors of the actual failure of the reaction pump. Summary of the Invention

[0006] The purpose of this invention is to provide a method for extracting fault features of industrial pumps based on time-series data analysis. This method selects candidate abrupt change moments that can reflect changes in system state, eliminates false abrupt changes caused by noise and transient disturbances, and accurately extracts feature vectors representing real faults from multi-dimensional time-series data of industrial pump operation.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a method for extracting industrial pump fault features based on time-series data analysis, comprising: The vibration acceleration signal and outlet pressure signal at multiple monitoring points during the operation of the industrial pump are acquired to form an original time-series data sequence. As a preferred embodiment of the present invention, the vibration acceleration signal and the outlet pressure signal are acquired synchronously under the same operating conditions of the industrial pump, and the sampling frequency is not less than 10 times the rotational frequency of the industrial pump to ensure the integrity of signal capture and the distinguishability of state changes.

[0008] Based on the vibration acceleration and outlet pressure values ​​at each sampling point in the original time-series data sequence, a two-dimensional state vector is constructed for each sampling point, and the two-dimensional state vectors of all sampling points are arranged in chronological order to form a state transition matrix. Preferably, when constructing the two-dimensional state vector, the vibration acceleration and outlet pressure values ​​at each sampling point are normalized to ensure that the normalized vibration acceleration and outlet pressure values ​​are within a preset numerical range. The normalized vibration acceleration value is used as the first component of the two-dimensional state vector, and the normalized outlet pressure value is used as the second component, thus forming the two-dimensional state vector corresponding to each sampling point. Normalization eliminates dimensional differences, allowing vibration and pressure change information to be fused at a unified scale, improving the stability of subsequent analysis.

[0009] Singular value decomposition (SVD) is performed on the state transition matrix to obtain a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix. The principal singular value sequence is then extracted from the singular value diagonal matrix. Preferably, when extracting the principal singular value sequence, the singular values ​​on the diagonal of the singular value diagonal matrix are arranged in descending order to obtain a descending singular value sequence. The third-largest number of singular values ​​from this descending singular value sequence is selected to form the principal singular value sequence, where the third-largest number is less than the rank of the singular value diagonal matrix. Retaining the principal singular values ​​can highlight fault-related abrupt changes in the system's dynamic characteristics while removing noise and weak interference components.

[0010] A difference operation is performed on the principal singular value sequence to obtain a difference singular value sequence. The time corresponding to the difference singular value in the difference singular value sequence where the numerical sign changes is marked as a candidate mutation time. Preferably, when performing the difference operation, for each principal singular value in the principal singular value sequence except the last principal singular value, the adjacent next principal singular value is subtracted from the principal singular value to obtain the difference singular value corresponding to the principal singular value. All difference singular values ​​corresponding to the principal singular values ​​are arranged in the original order of the principal singular value sequence to form the difference singular value sequence. When marking candidate mutation moments, starting from the first differential singular value in the differential singular value sequence, the current differential singular value is compared with the next differential singular value sequentially. If the product of the current differential singular value and the next differential singular value is less than zero, the moment corresponding to the next differential singular value is marked as a candidate mutation moment. If the product is equal to zero, the current differential singular value is skipped, and the comparison continues with the next and the following differential singular values. By utilizing the sign change of the differential singular values, the turning point in the evolution trend of the main singular value caused by the fault can be sensitively captured, achieving preliminary location of potential fault moments.

[0011] Centered on each candidate mutation moment, a local state sub-matrix within a preset time window is extracted from the state transition matrix. The rank of the local state sub-matrix is ​​calculated, and the candidate mutation moment corresponding to the local state sub-matrix with a rank greater than a preset threshold is determined as the fault characteristic moment. Preferably, when extracting the local state sub-matrix, the row index position of each candidate mutation moment in the state transition matrix is ​​determined. Using this row index position as the center row, a first number of row vectors are selected upwards and a second number of row vectors are selected downwards in the state transition matrix. The center row, the first number of row vectors, and the second number of row vectors are recombine according to the original time order to form the local state sub-matrix corresponding to the candidate mutation moment. More preferably, the first number and the second number are equal, and the total duration of the preset time window is not less than one rotation cycle of the industrial pump, so as to fully capture the state change characteristics of the fault impact within the complete cycle. When calculating the rank of the local state submatrix, a row echelon transformation is performed on the local state submatrix corresponding to each candidate mutation moment to obtain a row echelon matrix. The number of non-zero rows in the row echelon matrix is ​​counted, and the number of non-zero rows is used as the rank of the local state submatrix. The rank of each local state submatrix is ​​compared with a preset threshold, and the candidate mutation moments corresponding to local state submatrixes with rank greater than the preset threshold are selected as fault feature moments. By verifying the rank of the local state submatrix, it is possible to distinguish between pseudo-mutations caused by short-term operating condition fluctuations and changes in system state dimension caused by real faults, significantly improving the accuracy and robustness of fault feature moment extraction.

[0012] The two-dimensional state vector corresponding to the fault feature moment in the state transition matrix is ​​used as the fault feature vector, and a fault feature set consisting of all fault feature vectors is output. Preferably, for each fault feature moment, the row vector corresponding to that fault feature moment is located in the state transition matrix, and the values ​​of the first and second components of that row vector are used as the first and second components of the fault feature vector, respectively. The timestamp of the fault feature moment is used as the label of the fault feature vector and associated with the first and second components of the fault feature vector for storage. The final set of fault feature vectors directly characterizes the vibration and outlet pressure coupling state of the industrial pump at the fault moment, providing compact and complete input features for subsequent fault diagnosis and classification, reducing redundant data and retaining key information for fault identification.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: By constructing a two-dimensional state vector from the vibration acceleration and outlet pressure values ​​at the same sampling point, and arranging them in chronological order to form a state transition matrix, the row vectors of the matrix simultaneously contain both pump dynamics and fluid excitation information. Singular value decomposition (SVD) is performed on this state transition matrix to extract the principal singular value sequence, obtaining the main components of state evolution from the energy distribution level. The magnitude of the singular values ​​is related to the activity level of the system state; when the pump's operating state undergoes a substantial change, the principal singular value sequence will show a corresponding turning point. Differential operations are performed on the principal singular value sequence to obtain the differential singular value sequence. The moment when the sign of the differential singular values ​​changes is used to mark the candidate abrupt change moment. The change in the differential sign can sensitively reflect the turning point of the main energy distribution structure in the state matrix, without omission due to fluctuations in the absolute amplitude of the signal. This effectively captures the subtle state transitions that occur in the multi-dimensional parameter co-variation of early industrial pump faults.

[0014] After obtaining candidate mutation moments, a local state submatrix within a preset time window is extracted from the state transition matrix centered on each candidate mutation moment, and the rank of this submatrix is ​​calculated. The rank of the state transition matrix reflects the correlation and independent activity dimension of the system's state variables within a certain time period. When an industrial pump truly fails, the internal dynamic constraints are broken, the system's degrees of freedom increase, and the rank of the local state submatrix increases significantly. If the candidate mutation moment is caused only by instantaneous sensor noise or random shocks, and the intrinsic constraints between state variables have not fundamentally changed, the rank of the local state submatrix remains at a low level. By comparing the rank of the local state submatrix with a preset threshold, the candidate mutation moment is determined as the fault characteristic moment only when the rank is greater than the preset threshold. This effectively separates the moment representing the true system fault, removes false feature points caused by random disturbances, and the final output fault feature vector accurately characterizes the multidimensional state information of the pump at the moment the true fault occurs. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a flowchart of a method for extracting fault features of industrial pumps based on time-series data analysis; Figure 2 This is a flowchart of the process for normalizing raw time-series data and constructing two-dimensional state vectors; Figure 3 This is a flowchart of the principal singular value sequence extraction process; Figure 4 This is a flowchart of the difference operation and candidate mutation time marking of the principal singular value sequence; Figure 5 This is a schematic diagram illustrating the descending singular value sequence and the determination of the third quantity; Figure 6 It is the time series variation curve of the differential singular values ​​of the principal singular value sequence. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] See Figure 1This invention provides a method for extracting fault features of industrial pumps based on time-series data analysis. The method includes: acquiring vibration acceleration signals and outlet pressure signals at multiple monitoring points of the industrial pump during its operation to form an original time-series data sequence; constructing a two-dimensional state vector corresponding to each sampling point based on the vibration acceleration value and outlet pressure value of each sampling point in the original time-series data sequence, and arranging the two-dimensional state vectors of all sampling points in chronological order to form a state transition matrix; performing singular value decomposition on the state transition matrix to obtain a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix, and extracting principal singular values ​​from the singular value diagonal matrix. The sequence is processed by performing a difference operation on the principal singular value sequence to obtain a difference singular value sequence. The time corresponding to the difference singular value in the difference singular value sequence where the numerical sign changes is marked as a candidate mutation time. Taking each candidate mutation time as the center, a local state sub-matrix within a preset time window is extracted from the state transition matrix. The rank of the local state sub-matrix is ​​calculated, and the candidate mutation time corresponding to the local state sub-matrix with a rank greater than a preset threshold is determined as the fault feature time. The two-dimensional state vector corresponding to the fault feature time in the state transition matrix is ​​used as the fault feature vector, and a fault feature set composed of all fault feature vectors is output.

[0019] Example 1: In specific implementation, please refer to Figure 2 The system acquires vibration acceleration signals and outlet pressure signals at multiple monitoring points during the operation of the industrial pump, forming a raw time-series data sequence. This sequence consists of multiple sampling points ordered by time, each corresponding to a timestamp and a synchronously acquired vibration acceleration value and outlet pressure value. For all sampling points in the raw time-series data sequence, the system first iterates through all vibration acceleration values ​​to find the maximum and minimum values, and simultaneously iterates through all outlet pressure values ​​to find the maximum and minimum values.

[0020] The preset numerical range for normalization is [0,1]. A linear mapping method is used, and the normalization formula is expressed as: in, This represents the original signal value to be normalized. Indicates and The minimum value of the corresponding signal of the same type across all sampling points. Indicates and The maximum value of the corresponding signal of the same type across all sampling points. This represents the signal value obtained after normalization. The value range is [0, 1]. When and When they are equal, set directly. It is 0.

[0021] For each sampling point, the vibration acceleration value of that sampling point is taken as... The minimum vibration acceleration is taken as The maximum vibration acceleration is taken as Substituting these values ​​into the normalization formula above, the normalized vibration acceleration value is calculated; simultaneously, the outlet pressure value at this sampling point is used as... The minimum export pressure is taken as The maximum export pressure is taken as Substituting into the above normalization formula, the normalized outlet pressure value is calculated.

[0022] For each sampling point, after obtaining the normalized vibration acceleration value and the normalized outlet pressure value, the normalized vibration acceleration value is used as the first component of the two-dimensional state vector, and the normalized outlet pressure value is used as the second component of the two-dimensional state vector, thus forming a two-dimensional state vector corresponding to that sampling point. The above operation is performed on each sampling point in the original time series data sequence to form the two-dimensional state vector corresponding to each sampling point.

[0023] Example 2: In specific implementation, please refer to Figure 3 After performing singular value decomposition on the state transition matrix, we obtain the left singular vector matrix, the singular value diagonal matrix, and the right singular vector matrix. The singular value diagonal matrix is ​​a matrix where non-zero elements exist only on the main diagonal, and all other elements are zero. The process of extracting the principal singular value sequence starts directly from the singular value diagonal matrix.

[0024] Read all elements on the main diagonal of the singular value diagonal matrix one by one to obtain a set of singular values. Arrange all the singular values ​​in this set of singular values ​​in descending order to form a descending singular value sequence. The first singular value in the descending singular value sequence corresponds to the largest singular value, and the last singular value corresponds to the smallest singular value.

[0025] A positive integer parameter, denoted as the third quantity, is pre-defined. The specific value of the third quantity is determined by analyzing the decay characteristics of singular values ​​in the descending singular value sequence. The ratio of any two adjacent singular values ​​in the descending singular value sequence is calculated sequentially. When the ratio of the first singular value to the second singular value is greater than the ratio of the first singular value to a certain singular value, the position where the ratio jumps significantly is found. The number of singular values ​​before the significant jump is determined as the third quantity, while ensuring that the value of the third quantity is always less than the rank of the singular value diagonal matrix.

[0026] After determining the third number, select the first to the third number of singular values ​​from the descending singular value sequence, for a total of the third number of singular values. Then, arrange these third number of singular values ​​according to their order in the descending singular value sequence to form the principal singular value sequence. This completes the operation of extracting the principal singular value sequence from the singular value diagonal matrix.

[0027] In practice, the trend and fluctuation components of the main singular value sequence are separated. The smoothing scale is determined based on the real-time rotation period of the industrial pump. The rotation period is calculated from the real-time rotation speed collected by the speed sensor, with the unit being seconds, and the conversion relationship is that the rotation period equals sixty divided by the rotation speed value. Simultaneously, the number of sampling points corresponding to the rotation period is calculated using the sampling frequency set during signal acquisition. The number of sampling points corresponding to the rotation period is equal to the product of the rotation period and the sampling frequency, and the product result is rounded up as the lower bound of the local weighted regression smoothing window length. The local weighted regression smoothing uses a weighted local polynomial fitting to perform one iteration on each data point of the main singular value sequence. Each fitting local window covers all main singular value data points within half the window width before and after the center point. The weight of each data point within the window is generated by a cubic weight function based on the distance between the data point and the center point. The fitting polynomial is set to a quadratic polynomial. After fitting, the fitted value at the center point is taken as the smoothed value of the trend component sequence at the corresponding time. After traversing all time points of the main singular value sequence through the sliding window, a trend component sequence of the same length as the main singular value sequence is obtained. Then, the value of each sampling time of the main singular value sequence is subtracted from the value of the same time in the trend component sequence, and the result of the subtraction constitutes the fluctuation component sequence.

[0028] In some embodiments, the window length for local weighted regression smoothing is the rounded-up value of the number of sampling points corresponding to the rotation period multiplied by 1.5, which effectively filters out fluctuations with time scales shorter than the rotation period while fully preserving the gradually changing profile across the rotation period. In some embodiments, after the initial fitting, local weighted regression calculates the residuals of each data point and adjusts the weights according to the residual magnitude for iterative weighted fitting, with the number of iterations fixed at three, thereby reducing the impact of isolated spikes in the principal singular value sequence on the morphology of the trend component sequence.

[0029] Optionally, the cubic weight function is expressed as weight equal to one minus the cube of the normalized distance, where the normalized distance is the absolute distance of the data point from the center point divided by the half-width of the window. If the normalized distance exceeds one, the weight is reset to zero. Optionally, the quadratic polynomial fitting includes constant, linear, and quadratic terms. Within the local window, the horizontal axis of the two-dimensional coordinate system represents the relative time sequence, and the vertical axis represents the values ​​of the principal singular values.

[0030] It is understandable that abrupt state changes induced by faults such as bearing spalling or impeller rubbing in industrial pumps typically occur within a few rotational cycles, while changes in operating conditions, such as inlet valve regulation or pipeline resistance drift, last for far longer than ten rotational cycles. Separating the trend from the fluctuation by using the rotational cycle as a boundary allows the trend component sequence to absorb the slow drift of the operating conditions. It is also understandable that the fluctuation component sequence obtained by subtraction essentially eliminates the interference of large-scale trends, and its numerical changes are mainly dominated by rapid state transitions associated with faults, providing an input sequence that only reflects rapid changes for subsequent difference operations.

[0031] In practice, the construction of the trend component sequence relies on the point-by-point smoothing calculation of the main singular value sequence. After each local fitting of a center point is completed, the window is moved to the next sampling point and the process is repeated until all moments of the main singular value sequence are covered, forming a continuous trend component sequence. Then, the corresponding positions of the main singular value sequence and the trend component sequence are subtracted point by point. The output fluctuation component sequence is a sequence that fluctuates around the zero value, which can be directly used for subsequent difference sign detection.

[0032] See Figure 5 , Figure 5 The horizontal axis represents the sequence number of the singular values, and the vertical axis represents the magnitude of the corresponding singular value. The curve depicts the numerical trend of the singular value sequence arranged in descending order. Starting from sequence number 1, the singular value gradually decreases from approximately 1.0, showing a rapid downward trend. As the sequence number increases, the singular value gradually approaches zero, reflecting the decay characteristic of singular values ​​in a descending singular value sequence.

[0033] Figure 5 The location of the "third number" is marked with a dashed line and indicated by text. This location corresponds to the point in the singular value sequence where a significant decay jump occurs, located at approximately index 20. This is a critical cutoff point determined by comparing the ratios of adjacent singular values, marking the boundary of the main singular value sequence consisting of the first singular value and the third number of singular values.

[0034] Example 3: In specific implementation, please refer to Figure 4 After the extraction of the principal singular value sequence is completed, the principal singular value sequence contains a third number of principal singular values ​​arranged in descending order. The difference operation is performed on each principal singular value in the principal singular value sequence except for the last principal singular value.

[0035] The position index in the principal singular value sequence is... The principal singular value is denoted as Index the location as The principal singular value is denoted as For the first singular value in the principal singular value sequence... Each principal singular value, calculate and The difference is used as the first... The principal singular values ​​correspond to the difference singular values. The formula for calculating the difference singular values ​​is expressed as: in, Indicates the first singular value in the principal singular value sequence. Individual singular values The value range is from 1 to All integers, This represents the total number of principal singular values ​​in the principal singular value sequence. Equal to the third quantity; Indicates the first singular value in the principal singular value sequence. +1 principal singular value, that is, with The next adjacent principal singular value; Indicates by The calculated first There are 1,000 singular differences. The above calculation is a traversal process. Take in sequence This continues until all the principal singular values ​​in the sequence except the last principal singular value have been subtracted from their respective next adjacent principal singular values.

[0036] Each difference singular value is calculated. At the same time, record The corresponding time. The time markers are determined as those involved in the calculation. The corresponding sampling time in the original time-series data sequence. All according to Arranged in ascending order, they form a sequence of difference singular values.

[0037] After the differential singular value sequence is generated, the labeling operation for candidate mutation times begins. This is done by iterating through the sequence starting from the first differential singular value. During this iteration, the positions in the differential singular value sequence are indexed as follows: The difference singular value is denoted as the current difference singular value, and the position is indexed as... The difference singular value is recorded as the next difference singular value, and the traversal pointers are sequentially... Move from 1 to ,in This represents the total number of difference singular values ​​in the difference singular value sequence. equal .

[0038] In each comparison, the product of the current singular difference value and the next singular difference value is calculated. If the product is less than zero, the current singular difference value and the next singular difference value are considered to have opposite signs, and the time corresponding to the next singular difference value is marked as a candidate mutation time. If the product is equal to zero, the position of the current singular difference value is skipped, and the traversal pointer is moved to the next position. The value is incremented by 1, and the product of the new current difference singular value and its corresponding next difference singular value is compared. If the product value is greater than zero, the current difference singular value and the next difference singular value have the same sign, no candidate mutation time marker is generated, and the traversal pointer is incremented. The value increases by 1, and the process moves on to the next comparison.

[0039] Repeat the above comparison process until all adjacent differential singular value pairs in the differential singular value sequence have completed the product judgment, and finally output all marked candidate mutation times.

[0040] See Figure 6 , Figure 6 The horizontal axis represents time, ranging from 0 to 500, and the vertical axis represents the numerical value of the difference singular value, ranging from 0 to 0.05. The curve is the difference singular value sequence described in Example 3, reflecting the trend of the difference between adjacent singular values ​​in the main singular value sequence changing with time.

[0041] from Figure 6 It was observed that the differential singular value sequence exhibited significant fluctuations in the early stages (approximately 0 to 100), with the differential singular value values ​​repeatedly reaching peaks above 0.04 and frequently alternating between positive and negative values. This reflects the significant fluctuations in the singular value differences within the main singular value sequence, indicating that there may be a sudden change signal during this period, corresponding to an important change in the operating status of the industrial pump.

[0042] As time increases (approximately 100 to 300), the amplitude of the differential singular values ​​gradually decreases, the fluctuation frequency decreases, and the overall value approaches zero, indicating that the differences in the main singular value sequence weaken and the state of the industrial pump tends to be relatively stable.

[0043] In the later period (approximately 300 to 500), the difference singular values ​​remain at extremely low levels with minimal fluctuations, approaching zero. This indicates that the main singular value sequence changes very smoothly during this period, and the industrial pump operates stably without any significant abrupt changes.

[0044] Example 4: In practice, after all candidate mutation times are marked, a local state submatrix truncation operation is performed for each candidate mutation time. The state transition matrix is ​​composed of two-dimensional state vectors of all sampling points arranged in chronological order. Each row in the state transition matrix corresponds to a two-dimensional state vector at a sampling time, and the row index position corresponds one-to-one with the chronological order of the sampling times.

[0045] For a given candidate mutation time, first locate the row index position corresponding to the candidate mutation time in the state transition matrix, and denote this row index position as the center row. The total number of rows in the state transition matrix is ​​denoted as... The row index position of the center row is denoted as , For a value ranging from 1 to Integers between [a certain range].

[0046] A first quantity is preset, which is a positive integer. The value of the first quantity depends on the relationship between the preset time window length and the sampling period. Let the total duration of the preset time window be... The sampling period is denoted as The first quantity is determined by the following formula: in, Indicates the first quantity; This indicates the total duration of the preset time window. The value of is not less than one rotation cycle of the industrial pump. One rotation cycle of the industrial pump is calculated based on the real-time speed of the industrial pump. The speed is measured in revolutions per minute, and the rotation cycle is the reciprocal of the speed multiplied by 60, measured in seconds. This represents the sampling period, which is the time interval between two adjacent sampling points. Based on the sampling frequency calculation, the sampling frequency is set to be no less than 10 times the industrial pump rotation frequency, which is equal to the rotation speed divided by 60. The floor symbol indicates the rounding down function, used to ensure... The first quantity is an integer. The second quantity is set to be exactly equal to the first quantity.

[0047] Determine the row index position of the center row. and the first quantity Next, select row vectors upwards from the state transition matrix. The range for selecting row vectors upwards is from the row index position. To row index position ,common There are row vectors. Boundary constraints are applied to the selected upward range if... If the value is less than 1, the starting row index position for upward selection is corrected to 1. In this case, the actual number of row vectors selected upward is... indivual.

[0048] Select row vectors downwards from the state transition matrix. The range for selecting row vectors downwards is from the row index position. To row index position ,common There are row vectors. Boundary constraints are applied to the range selected downwards, if... Greater than Then the index position of the terminating row selected downwards will be corrected to... At this point, the actual number of row vectors selected downwards is indivual.

[0049] Arrange all upward-selected row vectors, the center row vector, and all downward-selected row vectors in their original time order, maintaining the order of the sampling times corresponding to the row vectors from earliest to latest. The combined matrix is ​​the local state submatrix corresponding to the candidate mutation time. The total number of rows in the local state submatrix is ​​the sum of the actual number of upward-selected row vectors, one center row vector, and the actual number of downward-selected row vectors. The number of columns in the local state submatrix is ​​the same as the number of columns in the state transition matrix, both being two columns.

[0050] Example 5: In practice, the rank of the local state submatrix corresponding to each candidate mutation time is calculated. The local state submatrix is ​​a matrix with... A real matrix with 2 rows and 2 columns, where This represents the total number of rows in the local state submatrix. The value is equal to the sum of the number of row vectors actually selected upwards, one central row vector, and the number of row vectors actually selected downwards.

[0051] A row echelon transformation is performed on the local state submatrix using Gaussian elimination. Gaussian elimination applies a series of elementary row operations to the local state submatrix, including swapping any two rows, multiplying a row by a non-zero constant, and adding a row to another row by a multiple of the previous row. The transformation begins with the first column of the local state submatrix. A pivot is selected in each column, and all elements below the pivot in the same column are eliminated to zero. The row echelon transformation terminates when all columns have been processed, or when all elements in the remaining unprocessed rows are zero, resulting in a row echelon matrix.

[0052] Count the number of non-zero rows in a row echelon matrix. A non-zero row is defined as a row vector containing at least one element whose absolute value is greater than a preset minimum value. The row, preset small quantity The value is This is used in numerical computation to distinguish non-zero elements from minimal quantities caused by rounding errors. The number of non-zero rows obtained by statistics is taken as the rank of the local state submatrix, and the numerical value of the rank is denoted as . .

[0053] The preset threshold value is determined based on the total number of rows in the local state submatrix. Adaptive determination. The preset threshold is denoted as... , The calculation formula is expressed as: in, This indicates the preset scaling factor. The value is 0.5. The symbol represents the rounding up sign. This represents the total number of rows in the local state submatrix. The basis for setting the value to 0.5 is as follows: During the normal operation of an industrial pump, the correlation between two-dimensional state vectors is high, and the rank of the local state submatrix is ​​usually at a low level. Using half of the total number of rows as a threshold can distinguish between low-rank stationary states and high-rank abrupt states. When the rank exceeds half of the number of rows, it indicates that the joint change of vibration acceleration and outlet pressure within the time window exhibits strong independent component characteristics, which is consistent with the dynamic performance when a fault occurs.

[0054] The rank of each local state submatrix With preset threshold Perform a numerical comparison. If Greater than Then, the candidate mutation time corresponding to the local state submatrix is ​​selected as a fault characteristic time. If Less than or equal to If the candidate mutation time corresponding to the local state submatrix is ​​not selected as the fault characteristic time, it will be discarded.

[0055] After all fault characteristic moments are determined, a fault feature vector extraction operation is performed for each fault characteristic moment. The row vector corresponding to the fault characteristic moment is located in the state transition matrix. This row vector consists of the values ​​of a first component and a second component. The first component is the normalized vibration acceleration value, and the second component is the normalized outlet pressure value. The value of the first component in the row vector corresponding to the fault characteristic moment is taken as the first component of the fault feature vector, and the value of the second component is taken as the second component of the fault feature vector.

[0056] The timestamp of the fault feature moment is used as the label of the fault feature vector. The label, the first component, and the second component of the fault feature vector are stored together in association. The association storage method is to create a record in the data structure, which contains a timestamp field, a first component field, and a second component field. The values ​​of the three fields are written simultaneously to ensure that the correspondence is fixed.

[0057] The vibration acceleration signal and outlet pressure signal are acquired synchronously under the same operating conditions of the industrial pump. Synchronous acquisition is achieved through multiple analog input channels of the same data acquisition system driven by a unified sampling clock. The sampling frequency setting process is as follows: First, the industrial pump's rotational frequency is acquired. The industrial pump's rotational frequency is calculated from the real-time speed of the industrial pump, with the speed measured in revolutions per minute (rpm). The rotational frequency is equal to the speed divided by 60, measured in Hertz (Hz). Next, the lower limit of the sampling frequency is calculated, which is the industrial pump's rotational frequency multiplied by 10. Finally, a specific frequency value not lower than the lower limit is selected as the sampling frequency. Using a sampling frequency not less than 10 times the industrial pump's rotational frequency ensures that at least 10 sampling points are acquired within each rotation cycle of the industrial pump, thus fully preserving the dynamic details of the rotating machinery's operation in the time series.

[0058] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for extracting fault features of industrial pumps based on time-series data analysis, characterized in that, include: The vibration acceleration signal and outlet pressure signal at multiple monitoring points of the industrial pump during operation are acquired to form the original time-series data sequence. Based on the vibration acceleration value and outlet pressure value of each sampling point in the original time series data sequence, a two-dimensional state vector corresponding to each sampling point is constructed, and the two-dimensional state vectors of all sampling points are arranged in time order to form a state transition matrix. Singular value decomposition is performed on the state transition matrix to obtain a left singular vector matrix, a singular value diagonal matrix and a right singular vector matrix, and the principal singular value sequence is extracted from the singular value diagonal matrix. The main singular value sequence is separated into trend components and fluctuation components. The trend components caused by gradual changes in operating conditions are removed, and the fluctuation component sequence reflecting sudden changes in state is retained. Perform a difference operation on the fluctuation component sequence to obtain a difference singular value sequence, and mark the time corresponding to the difference singular value in the difference singular value sequence where the numerical sign changes as a candidate mutation time; Centered on each candidate mutation moment, a local state sub-matrix within a preset time window is extracted from the state transition matrix, the rank of the local state sub-matrix is ​​calculated, and the candidate mutation moment corresponding to the local state sub-matrix with a rank greater than a preset threshold is determined as the fault characteristic moment. The two-dimensional state vector corresponding to the fault feature moment in the state transition matrix is ​​used as the fault feature vector, and the fault feature set composed of all fault feature vectors is output.

2. The method for extracting industrial pump fault features based on time-series data analysis according to claim 1, characterized in that, Based on the vibration acceleration value and outlet pressure value of each sampling point in the original time-series data sequence, a two-dimensional state vector corresponding to each sampling point is constructed, specifically as follows: The vibration acceleration value and outlet pressure value at each sampling point are normalized so that the normalized vibration acceleration value and outlet pressure value are within the preset value range. The normalized vibration acceleration value is used as the first component of the two-dimensional state vector, and the normalized outlet pressure value is used as the second component of the two-dimensional state vector, thus forming the two-dimensional state vector corresponding to each sampling point.

3. The method for extracting industrial pump fault features based on time-series data analysis according to claim 1, characterized in that, Extracting the principal singular value sequence from the singular value diagonal matrix specifically involves: Arrange the singular values ​​on the diagonal of the singular value diagonal matrix in descending order to obtain a descending singular value sequence; The third number of singular values ​​is selected from the descending singular value sequence to form the principal singular value sequence, wherein the third number is less than the rank of the singular value diagonal matrix; The trend component and fluctuation component of the principal singular value sequence are separated as follows: Using the rotation period of the industrial pump as the reference window length, a sliding window local weighted regression smoothing is performed on the principal singular value sequence to obtain the trend component sequence; the trend component sequence is then subtracted from the principal singular value sequence to obtain the fluctuation component sequence. The rotation period is calculated based on the real-time rotation speed of the industrial pump, and the window length of the local weighted regression smoothing is not less than the number of sampling points corresponding to one rotation period, so as to ensure that the trend component only reflects the slow change process with a time scale greater than the rotation period, and does not include the rapid change with the same time scale as the fault mutation.

4. The method for extracting industrial pump fault features based on time-series data analysis according to claim 1, characterized in that, Performing a difference operation on the principal singular value sequence yields a difference singular value sequence, specifically: For each principal singular value in the principal singular value sequence except the last principal singular value, subtract the next adjacent principal singular value from the principal singular value to obtain the difference singular value corresponding to the principal singular value. Arrange the difference singular values ​​corresponding to all principal singular values ​​in the original order of the principal singular value sequence to form the difference singular value sequence.

5. The method for extracting industrial pump fault features based on time-series data analysis according to claim 1, characterized in that, The time corresponding to the change of the sign of the difference singular value in the difference singular value sequence is marked as the candidate mutation time, specifically: Starting from the first difference singular value in the difference singular value sequence, the current difference singular value is compared with the next difference singular value in turn. If the product of the current difference singular value and the next difference singular value is less than zero, then the time corresponding to the next difference singular value is marked as a candidate mutation time. If the product of the current difference singular value and the next difference singular value is equal to zero, then skip the current difference singular value and continue to compare the next difference singular value with the one after that.

6. The method for extracting industrial pump fault features based on time-series data analysis according to claim 1, characterized in that, Centered on each candidate mutation moment, a local state submatrix within a preset time window is extracted from the state transition matrix, specifically as follows: Determine the row index position of each candidate mutation time in the state transition matrix, and take that row index position as the center row. Select a first number of row vectors upward and a second number of row vectors downward in the state transition matrix. The central row, the first number of row vectors, and the second number of row vectors are recombined in their original time order to form the local state submatrix corresponding to the candidate mutation moment.

7. The method for extracting industrial pump fault features based on time-series data analysis according to claim 6, characterized in that, The first quantity is equal to the second quantity, and the total duration of the preset time window is not less than one rotation cycle of the industrial pump.

8. The method for extracting industrial pump fault features based on time-series data analysis according to claim 1, characterized in that, Calculate the rank of the local state submatrix, and determine the candidate mutation time corresponding to the local state submatrix with a rank greater than a preset threshold as the fault characteristic time, specifically: Perform row echelon transformation on the local state submatrix corresponding to each candidate mutation time to obtain a row echelon matrix, and count the number of non-zero rows in the row echelon matrix, and use the number of non-zero rows as the rank of the local state submatrix; The rank of each local state submatrix is ​​compared with a preset threshold, and the candidate mutation time corresponding to the local state submatrix with a rank greater than the preset threshold is selected as the fault feature time.

9. The method for extracting industrial pump fault features based on time-series data analysis according to claim 1, characterized in that, The two-dimensional state vector corresponding to the fault feature moment in the state transition matrix is ​​used as the fault feature vector, specifically: For each fault characteristic moment, locate the row vector corresponding to the fault characteristic moment in the state transition matrix, and use the values ​​of the first component and the second component of the row vector as the first component and the second component of the fault characteristic vector, respectively. The timestamp of the fault feature moment is used as the label of the fault feature vector and associated with the first and second components of the fault feature vector for storage.

10. The method for extracting industrial pump fault features based on time-series data analysis according to claim 1, characterized in that, The vibration acceleration signal and the outlet pressure signal are collected synchronously under the same operating conditions of the industrial pump, and the sampling frequency is not less than 10 times the rotational frequency of the industrial pump.

Images