Centrifugal fan health state assessment method and system based on multi-scale operation fluctuation

Abstract

The invention discloses a centrifugal fan health state assessment method and system based on multi-scale operation fluctuation. Multi-source parameters such as vibration, temperature, current, pressure and flow are synchronously collected, statistical fluctuation characteristics, trend drift characteristics and frequency domain energy characteristics are extracted under different time scales, and a multi-scale operation fluctuation fingerprint vector is constructed; a health reference fingerprint database and a covariance matrix are established in the equipment health operation stage, the offset degree of the current state relative to the health reference is calculated through the mahalanobis distance, and continuous quantitative expression of the health state is achieved by constructing a health index; and further calculating a change rate based on the health index time sequence and constructing a trend prediction model to realize health state evolution trend analysis and risk advance early warning. According to the method, the normal working condition fluctuation and the progressive deterioration trend can be effectively distinguished, conversion from traditional fault judgment to continuous health assessment and trend prediction is achieved, and the precision and foresight of centrifugal fan operation state assessment are improved.

Description

Technical Field

[0001] This invention belongs to the field of centrifugal fan fault monitoring technology, and in particular relates to a method and system for assessing the health status of centrifugal fans based on multi-scale operational fluctuations. Background Technology

[0002] Centrifugal fans, as key power equipment widely used in industrial process systems, play a crucial role in conveying gases, maintaining pressure balance, and participating in heat exchange in fields such as power, metallurgy, petrochemicals, cement, environmental protection, and rail transportation. Their operating status directly affects the stability and safety of the entire production line. Once a fan experiences performance degradation or sudden failure, it often leads to abnormally high system energy consumption, increased vibration, excessive bearing temperature rise, and even equipment shutdown due to interlocking mechanisms, resulting in significant economic losses. Therefore, accurate, timely, and effective health status assessment of centrifugal fans is a vital foundation for achieving predictive maintenance and ensuring the reliable operation of industrial systems.

[0003] In existing technologies, centrifugal fan condition monitoring typically focuses on operating parameters such as vibration, bearing temperature, current, voltage, speed, flow rate, and pressure. Diagnosis is then performed by setting alarm thresholds or employing methods like spectrum analysis and envelope demodulation to analyze single or a limited number of characteristics. These methods often rely on exceeding set thresholds as the basis for judgment, or directly identify a fault type upon detecting a specific characteristic frequency. While this single-index, single-time-scale monitoring approach is effective in typical fault stages, it often only triggers alarms when the equipment exhibits obvious abnormalities or enters the early stages of failure, making it difficult to reflect the gradual degradation characteristics of the equipment over long-term operation.

[0004] Meanwhile, centrifugal fans often operate under varying loads, speeds, and complex environments in actual use, resulting in significant fluctuations and multi-scale variations in their operating parameters. For example, transient pressure fluctuations within the second range may originate from airflow pulsations or system disturbances, vibration amplitude changes within the minute range may be related to load switching, while trend drifts over hours or even longer periods may reflect chronic degradation processes such as bearing wear, impeller dust accumulation, or dynamic balance misalignment. Existing technologies often mix data from different time scales or only select fixed time windows for analysis, lacking a systematic characterization of multi-time-scale operational fluctuations, making it difficult to distinguish the essential differences between normal fluctuations and abnormal evolutions.

[0005] Furthermore, traditional fault diagnosis methods primarily focus on fault type identification, that is, determining the current fault state of equipment through feature matching given a known fault mode library. This method relies on the sufficient collection and labeling of typical fault samples. However, in real industrial environments, many degradation processes do not present as a clear, single fault mode, but rather as a sub-healthy state such as continuous deviations in performance indicators, decreased energy efficiency, or weakened stability. Existing technologies lack effective quantification methods for such non-faulty but deviated-from-optimal-operational-range states, often only identifying them after the equipment reaches a clear failure criterion, making it difficult to support refined operation and maintenance decisions.

[0006] At the data processing level, existing technologies typically rely on simple statistical analysis of real-time data or the construction of a single model for prediction, neglecting the overall operational characteristics of equipment at different stages of operation. Different wind turbines exhibit individual differences in their normal operation due to variations in structural dimensions, impeller type, installation conditions, and system compatibility. Evaluating them using only a uniform standard can easily lead to false alarms or missed alarms. How to dynamically compare and predict the trends of operational status while considering the individual characteristics of each piece of equipment remains a problem that current technologies urgently need to solve. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a method and system for assessing the health status of centrifugal fans based on multi-scale operational fluctuations. Specifically, the technical solution provided by this invention is as follows: A method for assessing the health status of centrifugal fans based on multi-scale operational fluctuations includes: S1. Collect multi-source operating parameters during the operation of the centrifugal fan. The operating parameters include at least vibration acceleration signal, bearing temperature signal, motor current signal, fan speed signal, inlet and outlet pressure signal, and flow signal. S2. Divide the collected operating parameter data into multiple preset time scales, and form a multi-scale data window using a sliding time window method under each time scale. S3. Within the time window of each time scale, extract fluctuation characteristics for each operating parameter. The fluctuation characteristics include at least statistical fluctuation characteristics, trend drift characteristics, and frequency domain energy characteristics. S4. The fluctuation features extracted from each operating parameter at the same time scale are concatenated in a preset order to form a feature vector at that time scale. The feature vectors at different time scales are then concatenated to construct a multi-scale operating fluctuation fingerprint vector. S5. When the centrifugal fan is in a healthy operating state, collect multi-scale operating fluctuation fingerprint vector samples, calculate the health benchmark fingerprint vector and the corresponding covariance matrix, and construct a health benchmark fingerprint library. S6. Match the multi-scale operation fluctuation fingerprint vector under the current operating state with the health benchmark fingerprint vector, and calculate the health offset based on the Mahalanobis distance of the covariance matrix. S7. Construct a health index based on the health offset, calculate the rate of change of the health index based on the time series of the health index, predict the development trend of the health status of the centrifugal fan, and output the health status assessment result or early warning information based on the prediction result.

[0008] Furthermore, in step S2: Let the first k The time length of each time scale is Then the first k The time window corresponding to each time scale is defined as:

[0009] in, Indicates the first k A sliding time window with multiple time scales; in actual operation, at least three time scales are set, namely short-time scale... Mesoscale and long-term scale And satisfy < < ; At each time scale, a sliding window approach is used to divide the data for each operating parameter, with overlap between adjacent time windows; within the time window... Within, any operating parameter signal Discretize into sequences ,in For the first in the window i Each sample value, The number of sampling points within the window corresponds to the sampling time point. N k ,and ,in, fs This refers to the sampling frequency of the corresponding operating parameters.

[0010] Furthermore, in step S3: The statistical fluctuation characteristics include the window mean. Standard deviation and volatility coefficient : , ,

[0011] in, Indicates the first kThe average value of the operating parameters within a time scale window reflects the steady-state center level at that time scale; This indicates the fluctuation range at that scale. Used to indicate the degree of relative fluctuation.

[0012] Furthermore, the trend drift feature described in S3 is used to identify whether a persistent trend of change exists within the window: A linear model is established by performing least squares linear fitting on the data within the time window. ,in, For the first k The trend slope at each time scale represents the rate of change of the operating parameter within that time scale. For intercept term; trend slope Based on the sampling time point using the least squares method With corresponding sampled values Calculation yielded:

[0013] Trend slope The sign of the value indicates whether the parameter is increasing or decreasing within that time scale, and its absolute value indicates the rate of change.

[0014] Furthermore, the frequency domain energy characteristics described in step S3 are: For vibration acceleration signals During the time window Perform a Fast Fourier Transform within the range to obtain the frequency domain representation. , For frequency f Complex spectral values ​​at that location, The amplitude spectrum is used to extract the dominant frequency amplitude. ,in The dominant frequency with the largest amplitude; extract the subharmonic amplitude. And calculate the harmonic ratio. Simultaneously calculate the proportion of high-frequency energy:

[0015] in, For the first k The proportion of high-frequency energy over a given timescale; and These are the high-frequency start and stop frequencies, respectively. It is the Nyquist frequency.

[0016] Furthermore, in step S4: In the k Time windows corresponding to each time scale Within this process, statistical fluctuation features, trend drift features, and frequency domain energy features are extracted for each operating parameter. Then, the features of all operating parameters at the same time scale are concatenated in a preset order to form a feature vector for that time scale. :

[0017] in, Indicates the first m The operating parameter in the first k The feature components are defined at various time scales, where M is the total number of operating parameters; each feature component includes the operating parameter within a time window. mean within Standard deviation Volatility coefficient Trend slope , main frequency amplitude Harmonic ratio and the proportion of high-frequency energy At least a portion of it; Obtaining the feature vectors corresponding to each time scale Then, normalization is performed on each feature component; after normalization, the normalized feature vectors corresponding to each time scale are... Concatenate the data sequentially according to time scale to construct a multi-scale operational fluctuation fingerprint vector. , K This represents the total number of time scales.

[0018] Furthermore, in step S5: After the centrifugal fan has been commissioned and confirmed to be in a healthy operating state, multi-scale operational fluctuation fingerprint vector samples are continuously collected under multiple time windows. Let the first... n The multi-scale operational fluctuation fingerprint vector corresponding to each healthy sample is: ,in , For the number of samples in the healthy phase, The baseline health fingerprint vector is obtained by averaging the fingerprint vectors of all healthy samples element by element. :

[0019] Construct a covariance matrix based on all healthy samples. :

[0020] Through the health benchmark fingerprint vector and covariance matrix Jointly build a health benchmark fingerprint database.

[0021] Furthermore, in step S6: Obtain the current multi-scale operational fluctuation fingerprint vector within any current time window. F ; Simultaneously, obtain the health baseline fingerprint vector from the health baseline fingerprint database. ; Covariance matrix constructed based on health baseline stages Calculate health offset B :

[0022] in, Covariance matrix The inverse matrix of , where the superscript T denotes transpose.

[0023] Furthermore, in step S7: Building a health index Used to characterize the health status of the current operating state, where, This is the sensitivity adjustment coefficient, used to adjust the degree of response of the health index to the offset. As the sliding time window continues to advance, a health index time series is formed. The rate of change of the health index is calculated based on the time series of the health index, and its discrete difference expression is as follows:

[0024] This indicates the rate of change of the health index between two adjacent time windows. The window step time; A linear prediction model is constructed based on the current health index and the aforementioned trend rate to predict future time. Health Index at All Times :

[0025] in, To predict the time span; The smoothed rate of change And there are:

[0026] in, L To smooth the window length; Predicting future health index It compares the results with preset health thresholds and outputs health status assessment results or early warning information.

[0027] A centrifugal fan health status assessment system based on the above method, the system includes the following modules: The data acquisition and access module is used to collect multi-source operating parameter data during the operation of the centrifugal fan; The time synchronization and preprocessing module is connected to the data acquisition and access module and is used to perform time alignment, outlier processing and filtering preprocessing on the multi-source operating parameter data. A multi-scale window construction module, connected to the time synchronization and preprocessing module, is used to divide the running parameter data according to multiple preset time scales and generate a multi-scale data window in a sliding time window manner. A multi-scale fluctuation feature extraction module, connected to the multi-scale window construction module, is used to extract fluctuation features from the operating parameter data within a time window at each time scale. The running fluctuation fingerprint construction module is connected to the multi-scale fluctuation feature extraction module. It is used to splice the fluctuation features at the same time scale to form a time scale feature vector, and to splice the feature vectors at different time scales after normalization to form a multi-scale running fluctuation fingerprint vector. The health benchmark fingerprint library module is used to construct health benchmark fingerprint vectors and covariance matrices when the centrifugal fan is in a healthy operating state, thus forming a health benchmark fingerprint library. The multi-scale matching and health offset calculation module is connected to the running fluctuation fingerprint construction module and the health benchmark fingerprint library module. It is used to match the current multi-scale running fluctuation fingerprint vector with the health benchmark fingerprint vector and calculate the health offset based on the covariance matrix. The health index and trend assessment module is connected to the multi-scale matching and health offset calculation module. It is used to construct a health index based on the health offset and calculate the rate of change and development trend based on the health index time series. The results output module is connected to the health index and trend assessment module and is used to output health status assessment results and early warning information.

[0028] Compared with the prior art, the present invention has at least the following beneficial effects: First, this invention breaks through the traditional centrifugal fan condition monitoring method that relies on single-index threshold alarms. It no longer depends solely on whether a vibration amplitude or temperature rise exceeds limits as the sole criterion. Instead, it constructs a multi-scale operational fluctuation feature system, systematically fusing statistical fluctuation characteristics, trend drift characteristics, and frequency domain energy structure characteristics of the equipment at different time levels to form a structured operational fluctuation fingerprint. This method can simultaneously capture multi-level information such as second-level transient disturbances, minute-level load fluctuations, and hourly-level trend drifts. This transforms the equipment health status from a simplified binary judgment of normal / fault into a continuously measurable health offset, fundamentally improving the precision of condition identification.

[0029] Secondly, this invention avoids information distortion caused by the mixed calculation of features from different time scales by introducing a multi-scale sliding window processing mechanism. Short-term scales primarily reflect shocks and pneumatic fluctuations, medium-term scales reflect the fluctuation characteristics brought about by changes in operating conditions, and long-term scales reveal structural degradation trends. Through scale separation and fusion, normal operating condition fluctuations and abnormal evolution trends are mathematically distinguished, significantly improving the ability to identify gradual deterioration processes. This has a clear advantage for chronic problems such as bearing wear, impeller dust accumulation, or dynamic imbalance, enabling the identification of health status deviation trends before obvious fault characteristics develop.

[0030] Furthermore, this invention constructs a health benchmark fingerprint database and uses a covariance matrix to characterize the natural correlation structure between features under healthy conditions, upgrading health status assessment from single-point comparison to spatial distribution comparison. By calculating health offset using Mahalanobis distance, it considers not only the amplitude changes of each feature but also the correlation relationships between features, thereby effectively reducing the probability of false positives and false negatives. Compared to simple Euclidean distance or single-index threshold methods, this invention can adapt to individual differences in different devices, achieving personalized health modeling without relying on a large number of fault samples, and has stronger engineering adaptability.

[0031] Furthermore, this invention constructs a health index model through index mapping and further calculates the time derivative of the health index to achieve a quantitative expression of the rate of health state evolution. Through trend rate analysis and a linear prediction model, it can proactively estimate the health state over a future period, enabling early risk warning. Compared to traditional passive monitoring methods that only alert when anomalies occur, this invention can identify degradation trends in advance, providing a time window for operational and maintenance decisions, reducing the risk of sudden downtime, and improving equipment availability.

[0032] In summary, this invention achieves a technological upgrade from fault identification to health evolution assessment through multi-scale fluctuation modeling, health distribution space construction, and trend prediction mechanisms. This not only improves the accuracy and sensitivity of centrifugal fan health status assessment but also enhances predictive maintenance capabilities and risk warning capabilities. Attached Figure Description

[0033] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0034] Figure 1 This is a schematic diagram of the centrifugal fan health status assessment method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the centrifugal fan health status assessment system provided in an embodiment of the present invention. Detailed Implementation

[0035] 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, and not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort are all within the scope of protection of the present invention.

[0036] Example 1 This embodiment provides a method for assessing the health status of centrifugal fans based on multi-scale operational fluctuations.

[0037] like Figure 1 As shown, the method mainly includes the following steps: I. Multi-scale operational data acquisition Centrifugal fans are typical rotating machinery, and their health status exhibits fluctuations in parameters such as vibration, temperature, current, pressure, and flow rate. Therefore, this embodiment selects multi-source parameters that reflect mechanical, aerodynamic, and electrical conditions for simultaneous acquisition. The acquisition module can employ conventional industrial data acquisition methods, such as data aggregation based on PLCs, DCS, or industrial edge gateways. These methods are mature technologies in the field, and their specific hardware implementation structures will not be elaborated here.

[0038] For vibration signals, a piezoelectric accelerometer mounted on the outside of the bearing housing is preferred for acquisition, with the measurement direction including at least one of radial and axial directions. Let the vibration acceleration signal be... ,in t The time variable is represented in seconds. The vibration signal sampling frequency should be no less than 5 kHz to ensure that high-frequency information such as the bearing characteristic frequency and the impeller passing frequency can be captured; in practical implementation, a sampling frequency of 10 kHz or higher is preferred. The sampled raw signal is then converted from an A / D converter and sent to the data processing unit.

[0039] The bearing temperature signal is denoted as The temperature is measured in degrees Celsius and can be obtained using a PT100 resistance temperature detector (RTD) or thermocouple sensor. The sampling frequency is typically set to 1 Hz to 5 Hz, which is sufficient to reflect changes in temperature rise trends. The motor current signal is denoted as... The unit is amperes, which can be obtained through a Hall current sensor or transmitter, with a sampling frequency preferably of 5 Hz. The fan speed is denoted as... The unit is revolutions per minute (rpm), which can be obtained through an encoder or speed sensor. The inlet and outlet pressures are denoted as... and The unit is Pascal, and the flow rate signal is denoted as... The unit is cubic meters per hour, and the data can be collected using conventional pressure transmitters and flow meters.

[0040] All signals should be timestamped to ensure synchronization among multiple parameters. A unified time base can be achieved through an industrial clock synchronization protocol. Preliminary preprocessing of the acquired data includes removing obvious outliers and compensating for packet loss; these are standard signal preprocessing techniques in the field and can be implemented using moving average filtering or median filtering algorithms.

[0041] After completing continuous data acquisition, this embodiment divides the time axis into multiple scales to reflect the operational fluctuation characteristics at different time levels. Let the time variable be... t Then the original continuous data at any time can be represented as ,in This represents any of the above operating parameters. This embodiment defines multiple time-scale windows. Let the first... k The time length of each time scale is Its unit is seconds, then the first k Time scale window It can be defined as:

[0042] In practical implementation, three typical scales can be selected: short-time scale. Take 5 seconds, medium timescale Take 300 seconds (i.e., 5 minutes), long-term scale. Use 3600 seconds (i.e., 1 hour). This parameter can be adjusted according to actual working conditions, but in principle it should meet the following requirements. < < It covers three levels: transient disturbances, medium-term fluctuations, and long-term trends.

[0043] For continuously running data, a sliding window approach is used for processing at each time scale. Let the sliding step size be... Preferred selection The step size is 10% to 20%, for example, 1 second for short-term scales, 30 seconds for medium-term scales, and 5 minutes for long-term scales. A sliding window can be used to form continuous multi-scale sample sequences, ensuring the continuity of the evaluation results.

[0044] Within each time window, the discrete sampled data within the window is represented as:

[0045] in, For the first in the window i Each sample value, The number of sampling points within the window corresponds to the sampling time point. N k ,and ,in fs This refers to the sampling frequency of the corresponding parameter. For example, if the vibration sampling frequency is 10 kHz, the short-time window... If the interval is 5 seconds, then N1 = 5 × 10000 = 50000 data points.

[0046] Using the above method, a synchronous, continuous, and structured multi-source operational data matrix at different time scales can be obtained. Let the multi-parameter set be... In the window Internally formed data matrix ,in m The number of parameters, for example, in this embodiment, includes 7 parameter variables: vibration acceleration, bearing temperature, motor current, fan speed, inlet and outlet pressure, and flow rate; N k This represents the number of sampling points within the window.

[0047] II. Multi-scale fluctuation feature extraction At any time scale k Let the continuous-time signal of a certain operating parameter be... This signal can represent vibration acceleration. Bearing temperature Motor current Import and export pressure , or traffic Any parameter. The signal is processed within a time window. Internal discretization yields the sequence. ,in .

[0048] First, extract the basic statistical fluctuation characteristics. The mean within the window is defined as:

[0049] Indicates the first k The average value of this parameter within a time scale window reflects the steady-state center level at that time scale.

[0050] After obtaining the mean, the standard deviation is calculated to characterize the intensity of fluctuations. The standard deviation is defined as:

[0051] in, This indicates the fluctuation range at that scale; This represents the deviation of each sample value from the mean; the standard deviation is obtained by taking the square root of the average of the sum of squares.

[0052] To eliminate the impact of differences in the dimensions of different parameters, a fluctuation coefficient (coefficient of variation) is introduced, defined as:

[0053] in This is a dimensionless parameter used to represent the degree of relative fluctuation. When When the value is close to zero, a very small positive number should be set. Make corrections, that is, use the denominator. To avoid numerical instability.

[0054] In addition to basic statistical features, this embodiment further extracts trend drift features to identify whether a persistent trend exists within the window. A least-squares linear fit is performed on the data within the window, and the fitting model is set as follows:

[0055] in, For the first k The trend slope over a time scale is expressed in parameter units per second. This is the intercept term. The solution is obtained using the least squares method, and its calculation formula is as follows:

[0056] Trend slope The sign of the value indicates whether the parameter is increasing or decreasing within that time scale, and its absolute value indicates the rate of change.

[0057] For vibration signals, it is also necessary to extract frequency domain fluctuation characteristics. Let's assume a vibration acceleration signal... In the window The discrete sequence is obtained within. Performing a Fast Fourier Transform (FFT) on the sequence yields its frequency domain representation:

[0058] in For frequency f Complex spectral values ​​at that location, This is the amplitude spectrum.

[0059] Extract the main frequency amplitude from the spectrum. This refers to the amplitude of the frequency component with the largest amplitude within the frequency range. Let this frequency be... ,but: It can also calculate the amplitude of subharmonics. Its corresponding frequency is If the frequency component exists in the spectrum, its amplitude is used; otherwise, the interpolated value from a nearby frequency is used. Calculate the subharmonic ratio:

[0060] This ratio reflects the degree of enhancement of nonlinear vibration.

[0061] In addition, to characterize the growth of high-frequency anomalous energy, the proportion of high-frequency energy is defined as follows:

[0062] in, For the first k The proportion of high-frequency energy over a given timescale; and These are the high-frequency start and stop frequencies, which can be set according to the device structure, for example, from 1 kHz to 5 kHz; The Nyquist frequency of the sampling system, i.e. The numerator represents the high-frequency energy, and the denominator represents the full-frequency energy. In actual calculations, discrete summation is used instead of integration.

[0063] After extracting the above statistical features, trend features, and frequency domain features, in the... k At least one feature vector is formed at each time scale:

[0064] The same processing is applied to all time scales, ultimately forming a multi-scale fluctuation feature set:

[0065] in K The quantity is the time scale.

[0066] III. Constructing the operational fluctuation fingerprint vector In the k Time scale Below, for any operating parameter The following features have been obtained: window mean Standard deviation Coefficient of variation Trend slope , main frequency amplitude Harmonic ratio and the proportion of high-frequency energy It should be noted that the feature sets corresponding to different operating parameters can be different. For example, temperature signals do not involve frequency domain dominant frequency features, while vibration signals do include frequency domain features. Therefore, in practical implementation, a corresponding feature set can be defined for each type of parameter, and then a unified vector can be formed by concatenating them. Those skilled in the art can flexibly combine the selected parameters, but it is essential to ensure that the order of the features remains consistent throughout the entire system.

[0067] In a single time scalek Next, the features of all parameters within the same time window are concatenated in a fixed order to form a feature vector of that scale, denoted as:

[0068] in, For the first k A feature vector of operational fluctuations at various time scales; each element in the vector corresponds to the extracted statistical fluctuation intensity, relative fluctuation ratio, trend evolution rate, and frequency domain energy structure information. If multiple operational parameters exist, the above vector should be expanded into a multi-parameter concatenation form. For example, if there are M operational parameters, then:

[0069] in Indicates the first m The operating parameter in the first k Feature vectors at various time scales. The concatenation process is simply an array join operation, a conventional data structure processing technique.

[0070] Because different characteristics have different units—for example, temperature is measured in degrees Celsius, vibration in m / s², and current in amperes—normalization is required before generating the final operational fingerprint. A baseline mean normalization method is preferred. Let's assume that the first value is obtained during the health baseline stage. k The first scale j The baseline mean of each feature is The baseline standard deviation is The normalized features are denoted as:

[0071] in For the current window k The first scale j One original feature value, These are the standardized feature values. The benchmark mean and benchmark standard deviation are obtained by continuously collecting data for no less than 30 days during the equipment's healthy operation phase, and calculating the statistical mean and standard deviation of the corresponding features over all time windows, which serve as the long-term health benchmark.

[0072] After standardization, the normalized feature vector is obtained. In the case of multiple time scales, feature vectors from all scales are concatenated in scale order to form a comprehensive operational fluctuation fingerprint vector.

[0073] in, For the first k Normalized feature vectors at each time scale; KThis represents the total number of time scales. For example, when K = 3, it corresponds to short-time, medium-time, and long-time scales, respectively.

[0074] This concatenation process creates a high-dimensional feature vector with dimensions of . , For the first k The number of feature dimensions at each time scale. This vector F It reflects the fluctuation structure information of the equipment at different time levels and is a complete fluctuation fingerprint of the current operating status.

[0075] IV. Establish a health benchmark fingerprint database After the equipment is installed and tested to confirm that it is in optimal operating condition, a health baseline fingerprint database needs to be established. The so-called optimal operating condition means that the equipment has no known mechanical faults, stable aerodynamic performance, vibration and temperature rise within the manufacturer's recommended range, and operating load fluctuates within the design range.

[0076] During the health baseline construction phase, operational data was continuously collected for no less than 30 days, and the data was processed using a sliding window method according to the aforementioned multi-scale partitioning method to obtain a series of multi-scale operational fluctuation fingerprint vectors. Let the... n The normalized multi-scale fingerprint vectors obtained under each time window are: ,in , This represents the total number of healthy phase windows. Each All are column vectors of dimension D, and their elements are consistent with the previous definition.

[0077] To obtain the health baseline fingerprint, this embodiment uses the statistical mean as the health center vector. The health baseline fingerprint vector is defined as:

[0078] This formula represents the average of all fingerprint vectors in the healthy phase to obtain the center position of healthy operation.

[0079] In obtaining the health baseline vector Next, to characterize the correlation structure among the features, it is necessary to construct the covariance matrix. Let the first feature be... i The eigencomponent in the th th is the ? n The values ​​in each sample are covariance matrix Defined as:

[0080] in, The covariance matrix is ​​D×D; Indicates the first n A column vector of differences between each sample vector and the health baseline vector; Transpose it; the outer product is a D×D matrix; sum over all samples and divide by . Obtain the sample covariance matrix. Elements of the covariance matrix. Indicates the first i The first feature and the second j The degree of linear correlation between features. The covariance matrix is ​​used to describe the joint distribution structure of each feature dimension under healthy operating conditions. Since the features have been standardized and their dimensions are unified, the numerical stability of the covariance matrix is ​​good.

[0081] Ultimately, the health benchmark fingerprint database consists of health benchmark vectors. and covariance matrix composition. Describing the central location of health status Describe the natural fluctuation structure between features in a healthy state.

[0082] V. Multi-scale matching and health offset calculation Within any given time window, the current operational fluctuation fingerprint vector has been obtained through multi-scale feature extraction and fusion steps. F To measure the deviation of the current operating state from the healthy state, this embodiment uses Mahalanobis distance as the matching metric. Mahalanobis distance can consider the correlation and scale differences between features, and is more suitable for multidimensional correlated data structures than simple Euclidean distance. The degree of health deviation B is defined as:

[0083] in, F The fluctuation fingerprint vector is used for the current time step; A health baseline fingerprint vector; Covariance matrix The inverse matrix; the final calculated result B is a non-negative scalar, representing the multidimensional weighted distance of the current state from the center of the healthy state.

[0084] It should be noted that, assuming the features have been standardized, the physical meaning of Mahalanobis distance can be understood as the degree of deviation of the current operating state from the standard deviation of the healthy feature distribution space. When B = 1, it indicates that the current state is approximately within one standard deviation of the healthy distribution; when B > 3, it indicates that the deviation has significantly increased.

[0085] To transform the offset into an intuitively understandable health index, the health index function HI is further constructed as follows:

[0086] in, This is the sensitivity adjustment coefficient, used to adjust the degree of response of the health index to the offset. It can be obtained through historical data calibration, for example, by regression analysis on samples with known mild and severe anomalies.

[0087] VI. Assessment of Health Development Trends After updating any sliding time window, the current health offset has been calculated. And obtain the health index HI based on the index mapping function. t ). HI( t This represents the quantified health status at the current moment, which can be used to form a discrete time series as time progresses.

[0088] in For the first i The time interval between the end times of the sliding windows is... The time interval is determined by the sliding window step size, for example, a short-time scale step size of 1 second, and a medium-time scale step size of 30 seconds or longer.

[0089] To characterize the evolutionary trend of health status, it is necessary to calculate the rate of change of health indices over time. The rate of change in continuous form is defined as:

[0090] in, This represents the first derivative of the health index with respect to time. Since the data in actual engineering is a discrete time series, the derivative is approximated using the difference form.

[0091] in, The current window's health index; This refers to the health index from the previous window; The window step time.

[0092] To reduce the impact of single-point fluctuations on trend judgment, a moving average smoothing process is preferred. Let the smoothing window length be... L (For example, taking 5 sampling points), the smoothed rate of change is defined as:

[0093] in, This is the smoothed trend rate, used for trend judgment.

[0094] The direction of the operational state can be determined by the sign of the trend rate. When When the health index remains stable within the statistical error range, the equipment is in a stable operating state; when When the health index continues to decline, it indicates that the operating status is gradually deviating from the healthy benchmark and showing a deteriorating trend; when When the health index rises, it may be due to the restoration of operating conditions or improved performance after maintenance.

[0095] To further achieve forward-looking assessment, this embodiment constructs a linear prediction model to estimate the health index at future times. Assuming that the health index changes approximately linearly over short time intervals, then in the future... The predicted value of the health index at any given time is defined as:

[0096] in, For predicting future health index, To predict the time span, you can set it as needed, for example, to 3600 seconds (1 hour) or 86400 seconds (1 day).

[0097] If the prediction result satisfies:

[0098] This allows for the issuance of early warning signals. Among them... The health threshold can be set based on historical fault data or experience, for example, 0.7 or 0.6.

[0099] It should be noted that the above linear prediction model is a first-order approximation model. Those skilled in the art can also use more complex time series prediction methods, such as AR models or Kalman filters, but these methods are all conventional prediction techniques. The core of this invention is to model the health index as a continuously evolving variable for trend analysis.

[0100] Through the above trend calculation and prediction process, a complete health evolution curve HI can be formed. t ), and the corresponding trend curve R( t This trend curve not only reflects the current degree of deviation, but also the rate of deviation, thus enabling the shift from state identification to state evolution prediction.

[0101] Example 2 Based on the above method, this embodiment provides a centrifugal fan health status assessment system based on multi-scale operational fluctuations. The system takes the synchronous acquisition of multi-source operating parameters as the entry point, the fluctuation feature modeling of multiple time scales as the core, and the health benchmark fingerprint database and matching calculation as the main line. Finally, it outputs continuous health offset, health index and its development trend, and gives early warning information when the preset threshold conditions are met, thereby realizing the transformation of centrifugal fan from fault point judgment to health evolution assessment.

[0102] like Figure 2As shown, the system includes a data acquisition and access module, a time synchronization and preprocessing module, a multi-scale window construction module, a multi-scale fluctuation feature extraction module, an operational fluctuation fingerprint construction module, a health benchmark fingerprint database construction and management module, a multi-scale matching and health offset calculation module, a health index and trend assessment module, and a result output and interaction module. The data acquisition and access module is used to acquire multi-source operating parameter data during the centrifugal fan's operation. These operating parameters include at least vibration acceleration, bearing temperature, motor current, fan speed, inlet and outlet pressure, and flow rate signals. Data acquisition can be achieved through PLC, DCS, or industrial edge gateways for data aggregation and forwarding. The time synchronization and preprocessing module is used to align the multi-source data with a unified timestamp to ensure that different parameters correspond under the same time benchmark. It also performs preprocessing operations such as outlier removal, packet loss compensation, and moving average filtering or median filtering on the acquired data to ensure stable and reliable subsequent calculation inputs.

[0103] The multi-scale window construction module is used to perform multi-timescale partitioning on the preprocessed continuous data stream and generate multi-scale data windows using a sliding window approach. This module incorporates multiple timescale parameters and sliding step size parameters, enabling the system to form window sequences covering different time levels, such as short-term, medium-term, and long-term, providing a data carrier for subsequent extraction of transient disturbances, medium-term fluctuations, and long-term drift. The multi-scale fluctuation feature extraction module calculates statistical fluctuation features, trend drift features, and frequency domain energy features corresponding to the vibration signal for each operating parameter within each timescale window: statistical fluctuation features characterize the average level and fluctuation intensity within the window; trend drift features obtain the trend slope through least-squares linear fitting to reflect the direction and rate of change; and frequency domain energy features extract the dominant frequency amplitude, harmonic ratio, and high-frequency energy proportion by performing a fast Fourier transform on the vibration signal to reflect changes in energy structure. The module outputs a structured feature set organized by parameters and timescale, providing input for fingerprint-based representation.

[0104] The fluctuation fingerprint construction module is used to concatenate the features of various parameters at the same time scale into a scale feature vector according to a preset order, and further normalize each feature to enable comparison of features with different dimensions in the same space. The benchmark mean and benchmark standard deviation used for normalization are derived from the statistical results of the health benchmark stage. The system ensures comparability between multiple runs and different windows through consistent feature order and consistent standardization rules. Based on this, the module concatenates the normalized feature vectors at different time scales in scale order to form a multi-scale operational fluctuation fingerprint vector with a unified dimension. This vector serves as the fluctuation fingerprint representing the current operational state.

[0105] The health baseline fingerprint database construction and management module is used to continuously collect fingerprint samples for a certain period of time after the centrifugal fan has been commissioned and confirmed to be in a healthy operating state. It then calculates the health baseline fingerprint vector and covariance matrix to form the health baseline fingerprint database. This module stores both the location of the health center and the natural correlation structure between features under healthy conditions. It also provides a versioned update mechanism, enabling the database to be rebuilt or updated when the equipment re-enters a stable health phase after component replacement or maintenance. This ensures that subsequent assessments still use the machine's health status as a reference, rather than applying a uniform threshold that could lead to false alarms or missed alarms.

[0106] The multi-scale matching and health offset calculation module matches the current running fingerprint with the health baseline fingerprint and calculates the health offset in Mahalanobis distance form based on the inverse of the covariance matrix. Through this matching calculation, the system outputs a continuous offset index, which quantifies the degree of deviation of the current state from the health distribution space, reflecting both whether a deviation has occurred and its magnitude. The health index and trend assessment module constructs a health index based on the health offset, forming a health index sequence that evolves over time. This module further performs difference calculations on the health index sequence to obtain the rate of change and smooths the trend rate using methods such as moving averages. Finally, it constructs a linear prediction model to estimate the health index for future times. When the predicted value is lower than a preset health threshold, this module sends an early warning trigger signal to the results output and interaction module, thereby achieving proactive risk assessment.

[0107] The above system can execute the centrifugal fan health status assessment method described in Embodiment 1, and has the corresponding functional modules and beneficial effects of the method. For technical details not described in detail in this embodiment, please refer to the centrifugal fan health status assessment method provided in Embodiment 1 of this invention.

[0108] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; under the concept of the present invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the present invention as described above, which are not provided in detail for the sake of brevity; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for health condition assessment of a centrifugal fan based on multi-scale running wave fluctuation, characterized in that, Comprise: S1, collect multi-source operating parameters in the running process of the centrifugal fan, the operating parameters at least including vibration acceleration signal, bearing temperature signal, motor current signal, fan speed signal, inlet and outlet pressure signal and flow signal; S2, the collected operating parameter data is divided according to a plurality of preset time scales, and a plurality of scale data windows are formed in a sliding time window manner under each time scale; S3, in each time scale time window, the fluctuation characteristics of each operating parameter are extracted, and the fluctuation characteristics at least include statistical fluctuation characteristics, trend drift characteristics and frequency domain energy characteristics; S4, the fluctuation characteristics of each operating parameter under the same time scale are spliced to form a characteristic vector of the time scale according to a preset order, and the characteristic vectors under different time scales are spliced to construct a multi-scale running fluctuation fingerprint vector; S5, in the healthy running state stage of the centrifugal fan, a plurality of scale running fluctuation fingerprint vector samples are collected, a healthy benchmark fingerprint vector and a corresponding covariance matrix are calculated, and a healthy benchmark fingerprint library is constructed; S6, the multi-scale running fluctuation fingerprint vector under the current running state is matched with the healthy benchmark fingerprint vector, and a health offset is calculated based on the Mahalanobis distance of the covariance matrix; S7, a health index is constructed according to the health offset, a health index change rate is calculated based on the time sequence of the health index, the development trend of the health state of the centrifugal fan is predicted, and a health state evaluation result or early warning information is output according to the prediction result.

2. The centrifugal fan health state evaluation method according to claim 1, wherein In step S2: Let the first k The time length of each time scale is Then the first k The time window corresponding to each time scale is defined as: wherein, represents a sliding time window of the k time scale; in actual operation, at least three time scales are set, respectively, a short time scale , a medium time scale , and a long time scale , and satisfy ;​​ At each time scale, a sliding window approach is used to divide the data for each operating parameter, with overlap between adjacent time windows; within the time window... Within, any operating parameter signal Discretize into sequences ,in For the first in the window i Each sample value, The number of sampling points within the window corresponds to the sampling time point. N k ,and ,in, fs This refers to the sampling frequency of the corresponding operating parameters.

3. The centrifugal fan health state evaluation method according to claim 2, wherein In step S3: The statistical fluctuation features include window mean , standard deviation and fluctuation coefficient : ， ， wherein, denotes the average value of the operating parameter within the k time scale window, reflecting the steady state central level at this time scale; denotes the fluctuation amplitude at this scale, to indicate the relative fluctuation degree.

4. The centrifugal fan health state evaluation method according to claim 3, wherein The trend drift characteristics in step S3 are used to identify whether there is a persistent change trend in the window: performing a least squares linear fit on the data within the time window to establish a linear model wherein, is a trend slope at a jth time scale, representing a rate of change of the operating parameter within the time scale, k is an intercept term; the trend slope is a trend slope at a jth time scale, representing a rate of change of the operating parameter within the time scale, is an intercept term; the trend slope is calculated from the sampling time points and the corresponding sampling values trend slope The positive and negative of the trend slope represent whether the parameter is rising or falling within the time scale, and the absolute value of the trend slope represents the speed of change.

5. The centrifugal fan health condition assessment method of claim 4, wherein, The frequency domain energy characteristics in step S3 are: For the vibration acceleration signal , a fast Fourier transform is performed in a time window to obtain a frequency domain expression , is a complex spectrum value at a frequency f , is an amplitude spectrum; a main frequency amplitude is extracted in the spectrum , is a main frequency frequency with the largest amplitude; a sub-harmonic amplitude is extracted , and a harmonic ratio value is calculated ; and a high frequency energy proportion is calculated simultaneously: wherein, is the high frequency energy ratio at the first k time scale; and are the high frequency start and end frequencies, respectively, is the Nyquist frequency.

6. The centrifugal fan health state assessment method according to claim 5, wherein In step S4: In the first k time scale corresponding to the time window , after extracting statistical fluctuation features, trend drift features and frequency energy features for each operating parameter, the features of all operating parameters under the same time scale are spliced in a predetermined order to form a feature vector under the time scale : in, Indicates the first m The operating parameter in the first k The feature components are defined at various time scales, where M is the total number of operating parameters; each feature component includes the operating parameter within a time window. mean within Standard deviation Volatility coefficient Trend slope , main frequency amplitude Harmonic ratio and the proportion of high-frequency energy At least a portion of it; After obtaining the feature vectors corresponding to each time scale After that, normalization processing is performed on each feature component; after the normalization processing is completed, the normalized feature vectors corresponding to each time scale are obtained The multi-scale running wave fluctuation fingerprint vectors are constructed by splicing in the order of time scales , K is the total number of time scales.

7. The centrifugal fan health condition assessment method of claim 6, wherein, In step S5: In the stage of confirming that the centrifugal fan is in a healthy running state through debugging, a plurality of multi-scale running fluctuation fingerprint vector samples in a plurality of time windows are continuously collected, and a plurality of multi-scale running fluctuation fingerprint vector samples in a plurality of time windows are continuously collected. n The multi-scale running fluctuation fingerprint vector corresponding to the first healthy sample is , is the number of healthy stage samples,​ averaging element-wise all healthy sample fingerprint vectors to obtain a healthy baseline fingerprint vector : Constructing a covariance matrix based on all healthy samples : through the health benchmark fingerprint vector and the covariance matrix together build a health benchmark fingerprint library.

8. The centrifugal fan health state assessment method according to claim 7, wherein In step S6: obtaining a current multi-scale running wave fluctuation fingerprint vector under any current time window F ; simultaneously obtaining a health benchmark fingerprint vector from a health benchmark fingerprint library ; Covariance matrix constructed based on health benchmark phase Computing health excursion B : wherein is the inverse of the covariance matrix , the superscript T denotes the transpose.

9. The centrifugal fan health condition assessment method of claim 8, wherein, In step S7: constructing a health index for characterizing the degree of health of the current operating state, wherein is a sensitivity adjustment factor for adjusting the degree of response of the health index to the offset; With the continuous advance of the sliding time window, a health index time series is formed A health index change rate is calculated based on the health index time series, and its discrete difference expression is represents the rate of change of the health index between the two adjacent time windows, is the window step time; constructing a linear prediction model based on the current health index and the trend rate, predicting a future time health index at the time : wherein is a prediction time span; is a smoothed rate of change and has: wherein, L is a smoothing window length; comparing the predicted future health index with the preset health threshold, outputting a health status assessment result or a warning information.

10. A centrifugal fan health condition assessment system based on the method of any one of claims 1 to 9, characterized by The system comprises: A data acquisition and access module for collecting multi-source operating parameter data in the running process of the centrifugal fan; A time synchronization and pretreatment module connected with the data acquisition and access module, for time alignment, outlier processing and filter pretreatment of the multi-source operating parameter data; A multi-scale window construction module connected with the time synchronization and pretreatment module, for dividing the operating parameter data according to a plurality of preset time scales, and generating a plurality of scale data windows in a sliding time window manner; A multi-scale fluctuation feature extraction module connected with the multi-scale window construction module, for extracting fluctuation characteristics of the operating parameter data in the time window of each time scale; A running fluctuation fingerprint construction module connected with the multi-scale fluctuation feature extraction module, for splicing the fluctuation characteristics under the same time scale to form a time scale characteristic vector, and splicing the characteristic vectors of different time scales after normalization to form a multi-scale running fluctuation fingerprint vector; A healthy benchmark fingerprint library module for constructing a healthy benchmark fingerprint vector and a covariance matrix in the healthy running state stage of the centrifugal fan, and forming a healthy benchmark fingerprint library; A multi-scale matching and health offset calculation module is connected with the running fluctuation fingerprint construction module and the health benchmark fingerprint library module, configured to match the current multi-scale running fluctuation fingerprint vector with the health benchmark fingerprint vector, and calculate a health offset based on the covariance matrix; A health index and trend evaluation module is connected with the multi-scale matching and health offset calculation module, configured to construct a health index according to the health offset, and calculate a change rate and a development trend based on a health index time sequence; A result output module is connected with the health index and trend evaluation module, configured to output a health state evaluation result and early warning information.

Images