Powder equipment vibration temperature fusion fault diagnosis method and system

Abstract

The invention discloses a vibration-temperature fusion fault diagnosis method and system for powder equipment, and belongs to the field of industrial equipment state monitoring and predictive maintenance, and the method comprises the steps: collecting a vibration acceleration and surface temperature sequence of an operation part of the powder equipment, calculating time delay through cross-correlation, executing resampling translation, and generating an alignment vibration and temperature sequence; variational mode decomposition and sliding window root mean square calculation are performed on the aligned vibration sequence, and a vibration energy flow sequence is constructed; training a multi-layer one-dimensional convolution kernel by using the vibration energy flow and the aligned temperature sequence, and generating a predicted temperature response sequence; a residual fluctuation variance is calculated, probability density distribution is constructed, and a dynamic confidence interval is generated; and performing online gradient updating on the convolution kernel by using the feature vector. According to the method, the space-time isomerism of multi-source data is eliminated, and adaptive fault early warning and model evolution based on a physical thermal mechanism are realized.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment condition monitoring and predictive maintenance, and in particular to a method and system for diagnosing vibration-temperature fusion faults in powder equipment. Background Technology

[0002] Powder processing equipment is widely used in industrial applications such as conveying, metering, mixing, classifying, and packaging powder materials. Its operating components typically require condition monitoring and maintenance management under continuous operation conditions. In current engineering practice, on-site monitoring often involves collecting and displaying operating parameters such as vibration, temperature, and rotational speed to assist in equipment inspection, maintenance scheduling, and anomaly handling.

[0003] Among related technologies, the invention patent application with publication number CN119225238A discloses an equipment monitoring visualization system based on digital twin technology. The system includes a data acquisition module, a virtual model creation module, a virtual-real mapping module, and an anomaly detection module.

[0004] The aforementioned technologies and similar application scenarios generally suffer from the following problems: First, monitoring data sources are scattered and data integration is insufficient, making it difficult to form a unified data view and consistent status representation; second, alarm methods based on single parameter thresholds are mainly triggered after the fact, which is difficult to meet the continuous operation and maintenance needs under complex working conditions; third, the diagnostic process relies heavily on human experience, resulting in insufficient consistency and reusability in handling. In addition, some digital twin visualization solutions suffer from insufficient fault prediction and diagnosis capabilities, as well as delays in the synchronous update of virtual models and physical devices. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method and system for fault diagnosis of powder equipment based on vibration and temperature fusion, which eliminates the spatiotemporal heterogeneity of multi-source data and realizes adaptive fault warning and model evolution based on physical thermal mechanisms.

[0006] The above objectives can be achieved through the following approach: A method for fault diagnosis of vibration-temperature fusion in powder processing equipment includes: collecting vibration acceleration sequences and surface temperature sequences of operating components of the powder processing equipment; performing cross-correlation calculations on the vibration acceleration sequences and surface temperature sequences, and recording the time delay of peak values; performing resampling and translation operations on the vibration acceleration sequences based on the time delay to generate aligned vibration sequences and aligned temperature sequences; performing variational mode decomposition and sliding window root mean square calculation on the aligned vibration sequences to obtain a multi-band vibration intensity matrix; performing weighted summation on the multi-band vibration intensity matrix using modal center frequencies to construct a vibration energy flow sequence; constructing a sliding window sample set using the vibration energy flow sequence and the aligned temperature sequence; initializing multi-layer one-dimensional convolutional kernels; and performing error backpropagation and parameter optimization on the sliding window sample set. The process involves iterative updates to generate a convolution kernel weight matrix. This matrix is ​​then used to perform a convolution operation on the vibration energy flow sequence to generate a predicted temperature response sequence. Pointwise difference and variance statistics are performed on the aligned temperature sequence and the predicted temperature response sequence to generate a residual fluctuation variance sequence. A probability density distribution function is constructed on the residual fluctuation variance sequence based on a kernel density estimation method. Integration and quantile calculations are then performed on the probability density distribution function to generate a dynamic confidence interval sequence. Topological violation feature vectors with amplitudes exceeding the dynamic confidence interval sequence are extracted from the residual fluctuation variance sequence. The Euclidean distance between the topological violation feature vectors and the cluster centers in a preset historical fault feature database is calculated. Finally, the topological violation feature vectors are used to perform gradient descent updates on the multi-layer one-dimensional convolution kernel.

[0007] Optionally, generating the aligned vibration sequence and aligned temperature sequence includes: performing mean reduction and fast Fourier transform on the vibration acceleration sequence and the surface temperature sequence to obtain a vibration spectrum sequence and a temperature spectrum sequence; calculating the conjugate product using the vibration spectrum sequence and the temperature spectrum sequence, and performing an inverse fast Fourier transform on the conjugate product to generate a cross-correlation function sequence; traversing the amplitude of the cross-correlation function sequence, extracting the x-coordinate index value of the point with the maximum amplitude, converting the x-coordinate index value into a time unit to obtain a time delay; performing linear interpolation resampling on the vibration acceleration sequence using the time delay to generate an intermediate sequence, and performing a reverse time displacement operation on the intermediate sequence to generate an aligned vibration sequence that is phase-synchronized with the surface temperature sequence, and using the surface temperature sequence to obtain the aligned temperature sequence.

[0008] Optionally, the construction of the vibration energy flow sequence includes: constructing a variational constraint optimization equation for the aligned vibration sequence; performing a frequency domain iterative search on the variational constraint optimization equation using a preset alternating direction multiplier algorithm to decompose and generate modal components; performing time-domain sliding window slicing on the modal components, calculating the root mean square amplitude in each slice, and arranging them according to the modal number and time dimension to generate a multi-band vibration intensity matrix; extracting the center frequency value for each modal component, constructing a frequency weighting vector, and performing matrix multiplication operation between the multi-band vibration intensity matrix and the frequency weighting vector to generate the vibration energy flow sequence.

[0009] Optionally, the decomposition to generate modal components includes: performing a fast Fourier transform on the aligned vibration sequence to generate a vibration frequency domain sequence; introducing Lagrange multiplier variables and using the vibration frequency domain sequence to construct an augmented Lagrange function containing modal spectrum variables and modal center frequency variables; performing iterative operations of the alternating direction multiplier algorithm on the augmented Lagrange function, updating the modal spectrum variables, the modal center frequency variables, and the Lagrange multiplier variables sequentially in a single iteration until the convergence condition is met, and outputting a converged modal spectrum sequence; and performing an inverse fast Fourier transform on the converged modal spectrum sequence to generate modal components.

[0010] Optionally, generating the predicted temperature response sequence includes: performing forward propagation calculations on the vibration energy flow sequence segments and aligned temperature sequence segments in the sliding window sample set; performing convolution operations on the vibration energy flow sequence segments using the multi-layer one-dimensional convolution kernel to generate a predicted intermediate temperature value; calculating the mean squared error loss function between the predicted intermediate temperature value and the aligned temperature sequence segment, and performing partial derivative calculations on the mean squared error loss function with respect to the multi-layer one-dimensional convolution kernel based on the chain rule to generate a weight gradient matrix; performing gradient descent updates on the multi-layer one-dimensional convolution kernel using the weight gradient matrix to generate a convolution kernel weight matrix; and performing sliding convolution point-by-point on the vibration energy flow sequence using the convolution kernel weight matrix to generate the predicted temperature response sequence.

[0011] Optionally, the method further includes: extracting the maximum amplitude value of the cross-correlation function sequence as a correlation strength scalar, and using the numerical result of the mean squared error loss function as a prediction error scalar; performing a ratio operation between the correlation strength scalar and the prediction error scalar to generate an adaptive update step size coefficient.

[0012] Optionally, generating the dynamic confidence interval sequence includes: performing point-by-point difference calculation on the aligned temperature sequence and the predicted temperature response sequence to generate a temperature residual sequence; calculating the variance of the temperature residual sequence using a sliding window to generate a residual fluctuation variance sequence; using the variance values ​​of the residual fluctuation variance sequence as input, performing kernel density estimation calculation using a preset Gaussian kernel function; calculating the weighted sum of the kernel function values ​​for each variance value point to generate a probability density distribution function; performing an integral operation on the probability density distribution function to obtain a cumulative distribution function; and solving for the lower quantile and upper quantile on the cumulative distribution function to generate the dynamic confidence interval sequence.

[0013] Optionally, generating the probability density distribution function includes: extracting variance values ​​using the residual fluctuation variance sequence, constructing a variance value sample set, and calculating the sample standard deviation of the variance value sample set; calculating a bandwidth parameter based on the sample standard deviation, and calculating a normalized difference between the variance value point and the sample value in the variance value sample set, substituting it into the Gaussian kernel function to obtain a kernel function value sequence; summing the kernel function value sequence to obtain a summation value, obtaining the number of samples using the variance value sample set, dividing the summation value by the product of the number of samples and the bandwidth parameter to obtain the density estimate of the variance value point, and constructing the probability density distribution function.

[0014] Optionally, the step of performing gradient descent update of the multi-layer one-dimensional convolutional kernel using the topological violation feature vector includes: performing interval comparison using the residual fluctuation variance sequence and the dynamic confidence interval sequence, marking the over-boundary index to calculate the over-boundary point count, the maximum length of the continuous over-boundary segment, and the maximum over-boundary amplitude, and constructing the topological violation feature vector; reading each cluster center vector from the historical fault feature database, calculating the Euclidean distance between the topological violation feature vector and the cluster center vector respectively, performing numerical multiplication of the adaptive update step size coefficient and the weight gradient matrix, and performing parameter update for the multi-layer one-dimensional convolutional kernel.

[0015] Based on the same inventive concept, this invention also provides a vibration-temperature fusion fault diagnosis system for powder equipment. The system includes: a vibration-temperature sequence acquisition and alignment module, used to acquire vibration acceleration sequences and surface temperature sequences of operating components of the powder equipment; perform cross-correlation calculations on the vibration acceleration sequences and surface temperature sequences, and record the time delay of peak values; and perform resampling and translation operations on the vibration acceleration sequences based on the time delay to generate aligned vibration sequences and aligned temperature sequences; a vibration energy flow construction module, used to perform variational mode decomposition and sliding window root mean square calculation on the aligned vibration sequences to obtain a multi-band vibration intensity matrix; and perform weighted summation on the multi-band vibration intensity matrix using modal center frequencies to construct a vibration energy flow sequence; and a thermal response convolution prediction module, used to construct a sliding window sample set using the vibration energy flow sequence and the aligned temperature sequence, and initialize multi-layer one-dimensional convolution kernels to predict the vibration energy flow sequence. The window sample set performs backpropagation of errors and iterative parameter updates to generate a convolution kernel weight matrix. This kernel weight matrix is ​​then used to perform convolution operations on the vibration energy flow sequence to generate a predicted temperature response sequence. A dynamic confidence boundary generation module performs point-by-point difference and variance statistics operations on the aligned temperature sequence and the predicted temperature response sequence to generate a residual fluctuation variance sequence. Based on a kernel density estimation method, a probability density distribution function is constructed on the residual fluctuation variance sequence, and integration and quantile solving are performed on the probability density distribution function to generate a dynamic confidence interval sequence. A fault identification and parameter update module extracts topological violation feature vectors from the residual fluctuation variance sequence whose amplitudes exceed the dynamic confidence interval sequence. It calculates the Euclidean distance between the topological violation feature vectors and the cluster centers in a preset historical fault feature library, and uses the topological violation feature vectors to perform gradient descent updates on the multi-layer one-dimensional convolution kernel.

[0016] Compared with the prior art, the present invention has the following advantages: 1. By introducing cross-correlation time delay analysis and resampling alignment mechanism, the problem of data spatiotemporal asynchrony between high-frequency vibration excitation and low-frequency temperature response in powder equipment is solved. The peak value of the cross-correlation function is used to accurately lock the time difference between frictional work and thermal energy accumulation, mapping heterogeneous data to the same physical phase. This eliminates interference caused by physical conduction lag, ensuring a strict correspondence between the input model data and the physical mechanism, thereby improving the effectiveness of multimodal data fusion and the accuracy of fault mapping. 2. A vibration energy flow sequence based on variational mode decomposition and frequency weighting was constructed, realizing a deep transformation from mechanical kinematic characteristics to thermodynamic energy characteristics. Unlike the black-box method that directly inputs the original signal, this method assigns higher energy weights to high-frequency friction components based on the law of frictional heat generation, quantifying the dynamic trend of mechanical work being converted into heat energy. This feature construction method driven by physical mechanisms not only enhances the interpretability of the model but also more keenly captures abnormal fluctuations in energy flow caused by early, subtle wear. 3. A strategy for dynamic confidence boundary generation and online adaptive update of model parameters based on residual kernel density estimation is proposed. By utilizing causal dilated convolution to learn the normal thermal response mechanism, a dynamic threshold fluctuating with operating conditions is generated through probability density integration, avoiding false alarms and false negatives associated with fixed thresholds. Simultaneously, online gradient updates are triggered using topological violation features, enabling the model to adjust weights in real time using fault samples. This provides the model with the ability to "self-evolve" as equipment ages and the environment changes, overcoming the pain point of performance degradation during long-term operation.

[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a method for diagnosing faults in the vibration-temperature fusion process of powder equipment according to an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the biaxial time-domain coupling characteristics of vibration energy flow and surface temperature response in an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of the temporal correlation between the feature distribution of the multi-band vibration intensity matrix and the adaptive update step size in an embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram of the structure of a powder equipment vibration temperature fusion fault diagnosis system according to an embodiment of the present invention. Detailed Implementation

[0023] 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.

[0024] Reference Figure 1 One embodiment of the present invention proposes a fault diagnosis method for vibration and temperature fusion in powder equipment, which eliminates the spatiotemporal heterogeneity of multi-source data and realizes adaptive fault warning and model evolution based on physical thermal mechanism.

[0025] The method described in this embodiment specifically includes: Vibration acceleration sequences and surface temperature sequences are collected from the operating components of the powder equipment. Cross-correlation calculations are performed on the vibration acceleration sequences and surface temperature sequences, and the time delay of the peak values ​​is recorded. Based on the time delay, resampling and translation operations are performed on the vibration acceleration sequences to generate aligned vibration sequences and aligned temperature sequences. Variational mode decomposition and sliding window root mean square calculation are performed on the aligned vibration sequence to obtain a multi-band vibration intensity matrix. The multi-band vibration intensity matrix is ​​then weighted and summed using the modal center frequencies to construct a vibration energy flow sequence. A sliding window sample set is constructed using the vibration energy flow sequence and the aligned temperature sequence. Multi-layer one-dimensional convolution kernels are initialized to perform error backpropagation and parameter iterative update on the sliding window sample set, generating a convolution kernel weight matrix. The convolution operation is then performed on the vibration energy flow sequence using the convolution kernel weight matrix to generate a predicted temperature response sequence. Perform point-by-point difference and variance statistics on the aligned temperature sequence and the predicted temperature response sequence to generate a residual fluctuation variance sequence. Construct a probability density distribution function on the residual fluctuation variance sequence based on the kernel density estimation method, and perform integration and quantile solving on the probability density distribution function to generate a dynamic confidence interval sequence. Extract the topological violation feature vectors from the residual fluctuation variance sequence whose amplitudes exceed the dynamic confidence interval sequence, calculate the Euclidean distance between the topological violation feature vectors and each cluster center in the preset historical fault feature library, and use the topological violation feature vectors to perform gradient descent update on the multilayer one-dimensional convolution kernel.

[0026] Optionally, generating the aligned vibration sequence and the aligned temperature sequence includes: The vibration acceleration sequence and the surface temperature sequence are subjected to mean reduction and fast Fourier transform to obtain the vibration spectrum sequence and the temperature spectrum sequence. For each sampling point of the vibration acceleration sequence, the arithmetic mean is calculated, and the arithmetic mean is subtracted from the value of each sampling point to obtain the mean-free vibration sequence. Similarly, for each sampling point of the surface temperature sequence, the arithmetic mean is calculated, and the arithmetic mean is subtracted from the value of each sampling point to obtain the mean-free temperature sequence. A Fast Fourier Transform (FFT) is performed on the mean-free vibration sequence to obtain the vibration spectrum sequence, and a Fast Fourier Transform (FFT) is performed on the mean-free temperature sequence to obtain the temperature spectrum sequence. The horizontal axis of the spectrum sequence represents the frequency index, and the vertical axis represents the complex spectral values.

[0027] The conjugate product is calculated using the vibration spectrum sequence and the temperature spectrum sequence, and an inverse fast Fourier transform is performed on the conjugate product to generate a sequence of cross-correlation functions. For each complex spectral value of the temperature spectrum sequence, take its conjugate complex value to obtain the temperature conjugate spectrum sequence. At the same frequency index, perform a complex multiplication operation on the complex spectral values ​​of the vibration spectrum sequence and the temperature conjugate spectrum sequence to obtain the conjugate product sequence. Perform an inverse fast Fourier transform on the conjugate product sequence to obtain the cross-correlation function sequence, where each sampling point of the cross-correlation function sequence is a complex value, and extract the magnitude of each sampling point as the corresponding amplitude.

[0028] For example, the sampling frequency of the vibration acceleration sequence is taken as... The sampling frequency of the surface temperature sequence is taken as The vibration and temperature sequences were recorded as equally spaced sampling points within the same monitoring period. After performing Fast Fourier Transform on the mean-reduced vibration and temperature sequences respectively to obtain the vibration spectrum sequence and temperature spectrum sequence, the temperature spectrum sequence was conjugate point by point and multiplied with the vibration spectrum sequence frequency by frequency index to form a conjugate product sequence. Then, the conjugate product sequence was inversely transformed by Fast Fourier Transform to obtain the cross-correlation function sequence, and the magnitude of the complex values ​​of each sampling point of the cross-correlation function sequence was taken to obtain the amplitude sequence.

[0029] Traverse the amplitude of the cross-correlation function sequence, extract the x-coordinate index value of the point with the maximum amplitude, convert the x-coordinate index value into a time unit, and obtain the time delay. Traverse the amplitude of the cross-correlation function sequence and lock the x-coordinate index value corresponding to the point with the maximum amplitude; read the sampling interval of the cross-correlation function sequence and perform a numerical multiplication operation between the sampling interval and the x-coordinate index value to obtain the time delay, where the sampling interval is the time interval between adjacent sampling points of the cross-correlation function sequence.

[0030] The vibration acceleration sequence is resampled by linear interpolation using the time delay to generate an intermediate sequence. The intermediate sequence is then subjected to a reverse time displacement operation to generate an aligned vibration sequence that is phase-synchronized with the surface temperature sequence. The aligned temperature sequence is then obtained using the surface temperature sequence.

[0031] The sampling interval of the surface temperature sequence is read as the target sampling interval. Linear interpolation resampling is performed on the vibration acceleration sequence to make the sampling interval of the resampled vibration sequence consistent with the target sampling interval, thus obtaining an intermediate sequence. The number of translation points is calculated based on the ratio of the time delay to the target sampling interval, and a reverse time displacement operation is performed on the intermediate sequence to obtain an aligned vibration sequence. The surface temperature sequence is then output as an aligned temperature sequence according to the time grid corresponding to the target sampling interval.

[0032] For example, the x-coordinate index of the point where the amplitude of the cross-correlation function sequence is maximum is taken as... The sampling interval of the cross-correlation function sequence is taken as... The delay is The target sampling interval for the surface temperature sequence is taken as... Then the number of translation points is rounded up to 1. Point, perform reverse time shift on the intermediate sequence The points were aligned to obtain the vibration sequence, and the surface temperature sequence was then sorted according to... The time raster output yields an aligned temperature sequence.

[0033] Optionally, the construction of the vibration energy flow sequence includes: A variational constraint optimization equation is constructed for the aligned vibration sequence. The variational constraint optimization equation is then subjected to frequency domain iterative search using a preset alternating direction multiplier algorithm to decompose and generate modal components. First, a Fast Fourier Transform (FFT) is performed on the aligned vibration sequence to obtain the vibration frequency domain sequence. Based on this, a variational constraint optimization equation is constructed. This equation solves for a set of modal frequency domain components with finite bandwidth and their corresponding center frequencies, minimizing the sum of the bandwidth metrics of each modal frequency domain component and satisfying the reconstruction constraint that the superposition of all modal frequency domain components in the frequency domain is consistent with the vibration frequency domain sequence. To solve this constraint optimization problem, the Alternating Direction Multiplier (ADMM) algorithm is employed. A quadratic penalty factor and Lagrange multiplier variables are introduced to transform the variational constraint optimization equation into an augmented Lagrange function. The quadratic penalty factor is used to adjust the penalty strength of the reconstruction constraint, and the convergence tolerance is used to limit the iteration stopping condition. Subsequently, the following three sub-steps are iteratively updated alternately in the frequency domain until the convergence condition is met: Keeping the center frequency variables and Lagrange multiplier variables of each mode unchanged, an update operation is performed on each mode frequency component based on the current reconstruction residual distribution in the frequency domain to obtain the frequency domain waveform of that mode in the corresponding frequency band; the current iterative value of each mode frequency component is obtained after the update; keeping the updated mode frequency components and Lagrange multiplier variables unchanged, the frequency centroid of the power spectrum distribution of each mode frequency component is calculated, and this frequency centroid is used as the updated center frequency value of that mode component, thereby achieving adaptive tracking of the center frequency to the frequency domain energy concentration; keeping the updated mode frequency components and center frequency unchanged, the reconstruction error between the vibration frequency domain sequence and the superposition result of all mode frequency components is calculated, and an update operation is performed on the Lagrange multiplier variables based on this reconstruction error to gradually satisfy the reconstruction constraints. The above three sub-steps are executed cyclically. Iteration stops when the change in the frequency domain component of each modality in two adjacent iterations is less than the convergence tolerance or the iteration limit is reached, and the converged set of frequency domain components is output. Finally, an inverse fast Fourier transform is performed on each converged frequency domain component to obtain the corresponding time domain modal component, while keeping the sampling point index of the modal component consistent with that of the aligned vibration sequence on the time axis.

[0034] The modal components are sliced ​​using a time-domain sliding window, the root mean square amplitude of each slice is calculated, and the slices are arranged according to the modal number and time dimension to generate a multi-band vibration intensity matrix. For each modal component, segmentation is performed along the time axis according to the window length, with adjacent segments sliding according to the window step size. The sampled values ​​within each segment are squared and their arithmetic mean is calculated. Then, the square root of the arithmetic mean is taken to obtain the root-mean-square (RMS) amplitude for that segment. The RMS amplitudes of each modal component obtained in each time segment are arranged in chronological order to obtain the corresponding RMS amplitude sequence. The RMS amplitude sequences corresponding to each modal component are then arranged by modal index to construct a multi-band vibration intensity matrix, where the row index corresponds to the modal index and the column index corresponds to the time segment index.

[0035] For example, the total number of sampling points for the aligned vibration sequence is taken as... Point, the number of modal components is taken as ; The window length is set to The step size of the point and window is taken as If a point is defined, then each modal component can form... The time segment; for the first time segment The first modal component Segmentation Square each point in turn, sum them, and divide by 1. The root mean square amplitude is obtained by taking the square root, and the rest are obtained in the same way. The root mean square magnitudes of each segment are taken and arranged into a root mean square magnitude sequence, and then... The root mean square amplitude sequences of each modal component are stacked according to their modal indices to obtain... OK The multi-band vibration intensity matrix of the column.

[0036] For each modal component, the center frequency value is extracted, a frequency weighting vector is constructed, and matrix multiplication is performed between the multi-band vibration intensity matrix and the frequency weighting vector to generate a vibration energy flow sequence.

[0037] For each modal component, a Fast Fourier Transform (FFT) is performed to obtain a modal spectrum sequence. The amplitudes of the modal spectrum sequence are iterated, and the frequency corresponding to the point of maximum amplitude is identified as the center frequency of that modal component. The center frequencies of each modal component are arranged according to their modal indices to construct a frequency weighting vector. Matrix multiplication is performed on the multi-band vibration intensity matrix and the frequency weighting vector to obtain the vibration energy flow sequence. The value of the vibration energy flow sequence at each time segment is the result of the inner product operation between the column vector of the multi-band vibration intensity matrix and the frequency weighting vector corresponding to that time segment. Through the above frequency weighting calculation, the vibration energy flow values ​​for each time slice can be obtained, such as... Figure 2 As shown in the figure, the gradient-colored bar chart shows the change in vibration energy flow amplitude after variational mode decomposition and frequency weighting, and the line chart with half-solid markings shows the surface temperature response curve after spatiotemporal alignment. The two show a physical causal correspondence in the time domain.

[0038] For example, The center frequencies of the modal components are taken sequentially as follows: And construct a frequency weighted vector according to the modal index, at the th At each time segment, the column vector of the multi-frequency band vibration intensity matrix is ​​taken as... Then, by multiplying the column vector element-by-element by the frequency weighted vector and summing the results, we obtain the value of the vibration energy flow sequence at that time segment. .

[0039] Optionally, the decomposition generates modal components including: Perform a fast Fourier transform on the aligned vibration sequence to generate a vibration frequency domain sequence; The discrete Fourier transform is calculated based on the sampling point sequence of the aligned vibration sequence to obtain the vibration frequency domain sequence. Each frequency index of the vibration frequency domain sequence corresponds to a complex spectral value, and the mapping relationship between the frequency index and the frequency value is recorded for subsequent extraction of the center frequency variable.

[0040] By introducing Lagrange multiplier variables, an augmented Lagrange function containing modal spectrum variables and modal center frequency variables is constructed using the vibration frequency domain sequence; The modal spectrum variables are defined as complex sequences of the same length as the vibration frequency domain sequence, and the modal center frequency variables are defined as a set of scalar values ​​of the corresponding frequency spectra of each modal spectrum variable. An augmented Lagrangian function is constructed based on the vibration frequency domain sequence, wherein the augmented Lagrangian function includes: a bandwidth metric term with the bandwidth of each modal spectrum variable in the neighborhood of its modal center frequency variable as the independent variable; a reconstruction difference term with the difference between the superposition of each modal spectrum variable and the vibration frequency domain sequence as the independent variable; and a constraint term with the reconstruction difference term as the independent variable and associated with the Lagrange multiplier variable.

[0041] For example, the number of modal components is taken as The upper limit of iteration is taken as The iteration stopping condition is that the root mean square value of the difference sequence of modal spectral variables between two adjacent iterations is less than 1. ; in the After the second iteration The modal center frequency variables are taken as follows: , No. After the next iteration, it is updated to... , No. The root mean square value of the time difference in the second iteration reaches And the stopping condition is met, the output includes A convergent modal spectrum sequence of modal spectral variables.

[0042] The augmented Lagrange function is subjected to the alternating direction multiplier algorithm iterative operation. In each iteration, the modal spectrum variable, the modal center frequency variable, and the Lagrange multiplier variable are updated sequentially until the convergence condition is met, and the converged modal spectrum sequence is output. Initialize the modal spectrum variables, modal center frequency variables, and Lagrange multiplier variables, and set the iteration stopping condition. In each iteration, keep the modal center frequency variables and Lagrange multiplier variables unchanged and update the modal spectrum variables. Then, keep the updated modal spectrum variables and Lagrange multiplier variables unchanged and update the modal center frequency variables again. Then, keep the updated modal spectrum variables and modal center frequency variables unchanged and update the Lagrange multiplier variables again. Calculate the root mean square value of the difference sequence of modal spectrum variables between two adjacent iterations. When the iteration stopping condition is met, output the convergent modal spectrum sequence, which is the set of all modal spectrum variables.

[0043] Perform an inverse fast Fourier transform on the convergent mode spectrum sequence to generate mode components.

[0044] Perform an inverse fast Fourier transform on each modal spectral variable in the convergent modal spectral sequence to obtain the corresponding modal component, while keeping the sampling point number of the modal component consistent with that of the aligned vibration sequence on the time axis.

[0045] For example, when the convergent mode spectrum sequence contains When there are multiple modal spectral variables, for Each modal spectral variable is obtained by performing an inverse fast Fourier transform. Each modal component, and The modal components are arranged by modal number and used for time-domain sliding window slicing and root mean square magnitude calculation.

[0046] Optionally, generating the predicted temperature response sequence includes: Forward propagation calculations are performed on the vibration energy flow sequence fragments and aligned temperature sequence fragments in the sliding window sample set. Convolution operations are performed on the vibration energy flow sequence fragments through the multi-layer one-dimensional convolution kernel to generate intermediate predicted temperature values. The sliding window sample set consists of multiple sample pairs, each containing a vibration energy flow sequence segment and an aligned temperature sequence segment at the same window location. For any vibration energy flow sequence segment, a convolution input vector is formed by sequentially taking the current sampling point and several previous sampling points in chronological order. The convolution input vector is then multiplied element-wise by the weight sequence of the first-layer one-dimensional convolution kernel and summed to obtain the corresponding sampling point values ​​of the first-layer convolution output sequence. This element-wise multiplication and summation operation is repeated along the time axis of the vibration energy flow sequence segment to obtain the first-layer convolution output sequence. The first-layer convolution output sequence is then used as the input to the second-layer one-dimensional convolution kernel, and the element-wise multiplication and summation operation is repeated to obtain the second-layer convolution output sequence. The second-layer convolution output sequence is then output as the predicted intermediate temperature value in chronological order of the sampling points.

[0047] Calculate the mean squared error loss function between the predicted temperature median and the aligned temperature sequence segment, and perform partial derivative calculation of the mean squared error loss function with respect to the multilayer one-dimensional convolution kernel based on the chain rule to generate a weight gradient matrix; For each pair of samples in the sliding window sample set, the difference between the predicted median temperature and the aligned temperature sequence segment at the same sampling point is calculated, and the difference is squared. The squared differences of each sampling point within the window are summed and divided by the number of sampling points within the window to obtain the mean squared error loss function value for that sample pair. The mean squared error loss function values ​​of multiple sample pairs in the sliding window sample set are summed and divided by the number of sample pairs to obtain the total mean squared error loss function value. Based on the chain rule, the partial derivatives of the total mean squared error loss function value with respect to the weight values ​​of the first-layer one-dimensional convolution kernel and the weight values ​​of the second-layer one-dimensional convolution kernel are calculated, and the partial derivatives are arranged according to the layer number and weight number to form the weight gradient matrix. The mean squared error loss function is shown below: , in, It is the mean squared error loss function, which is calculated from the sliding window sample set; It represents the number of sample pairs, which is the number of sample pairs within the sliding window sample set. This is the number of sampling points within the window, taking the number of sampling points for each vibration energy flow sequence segment and matching it with the aligned temperature sequence segment; It is the first The nth sample pair The aligned temperature values ​​of each sampling point are directly read from the aligned temperature sequence segment; It is the first The nth sample pair The predicted temperature for each sampling point is calculated using forward propagation. Physically, this means that within each sliding window, the difference between the median predicted temperature and the aligned temperature is squared and averaged within the window, then averaged over all sample pairs to obtain the mean squared deviation metric at the same time scale. The number of sampling points within the window is determined by the sliding window length, and the number of sample pairs is determined by the sliding window step size and the sequence length.

[0048] For example, the length of the sliding window is continuous. Each sampling point has a continuous sliding window step size. If there are 1 sampling point, then the number of sampling points within the window is 1. When the length of the vibration energy flow sequence corresponds to the formation At window positions, the number of sample pairs is For each window position, an aligned temperature sequence fragment is read to obtain an aligned temperature numerical sequence, and a predicted intermediate temperature value sequence is obtained through two layers of one-dimensional convolution operations; the two are then subtracted point by point and squared before... The mean squared error loss function value for the window is obtained by averaging over each sampling point, and then... The total value of the mean squared error loss function is obtained by averaging the mean squared error loss function values ​​of each window.

[0049] The weight gradient matrix is ​​used to perform gradient descent update on the multi-layer one-dimensional convolution kernel to generate a convolution kernel weight matrix. The convolution kernel weight matrix is ​​then used to perform point-by-point sliding convolution on the vibration energy flow sequence to generate a predicted temperature response sequence.

[0050] Update operations are performed on the weight values ​​of the first-layer one-dimensional convolutional kernel and the weight values ​​of the second-layer one-dimensional convolutional kernel. The update operation is to subtract the product of the learning rate and the corresponding partial derivative from the current weight value to obtain the convolutional kernel weight matrix. Using the convolutional kernel weight matrix, a sliding convolution is performed point-by-point on the vibration energy flow sequence. The element-wise product of the convolution input vector and the convolutional kernel weight sequence is calculated and summed point by point to obtain the values ​​of each sampling point of the predicted temperature response sequence. The predicted temperature response sequence is then output in chronological order.

[0051] For example, when a certain weight value of the first-layer one-dimensional convolution kernel is And the corresponding partial derivative values ​​are The learning rate value is When the update operation is performed, the updated weight values ​​are obtained as follows: The update operation is repeated for each weight value to obtain the convolution kernel weight matrix. Then, the vibration energy flow sequence at any sampling point and several sampling points before it are used to form the convolution input vector. This vector is multiplied element-wise with the updated convolution kernel weight sequence and summed to obtain the predicted temperature response sequence value for that sampling point. The calculation is repeated along the time axis to obtain the complete predicted temperature response sequence.

[0052] Optionally, the method further includes: The maximum amplitude value is extracted from the cross-correlation function sequence as a correlation strength scalar, and the numerical result of the mean squared error loss function is used as a prediction error scalar. Perform a traversal comparison of the amplitudes at each sampling point of the cross-correlation function sequence, lock the amplitude value of the point with the largest amplitude, and record this amplitude value as a correlation strength scalar. Calculate the total value of the mean squared error loss function for the sliding window sample set, and record the total value of the mean squared error loss function as a prediction error scalar.

[0053] For example, when the amplitude of the cross-correlation function sequence is at its maximum, the amplitude value is... At that time, Recorded as a scalar of correlation strength; when the total value of the mean squared error loss function is At that time, The record is a scalar of prediction error.

[0054] The correlation strength scalar and the prediction error scalar are used to perform a ratio calculation to generate an adaptive update step size coefficient.

[0055] The correlation strength scalar is used as the numerator and the prediction error scalar is used as the denominator to perform a numerical division operation to obtain the ratio result, and the ratio result is recorded as the adaptive update step size coefficient; when the prediction error scalar is zero, the minimum non-zero value of the mean square error loss function values ​​in the sliding window sample set is extracted as the denominator replacement value, and the numerical division operation is performed with the denominator replacement value.

[0056] For example, when the correlation strength scalar is And the prediction error scalar is When, execute get And record it as the adaptive update step size coefficient; when the prediction error scalar is And the denominator substitution value is When, execute get And record it as the adaptive update step size coefficient.

[0057] Optionally, generating the dynamic confidence interval sequence includes: Perform point-by-point difference operation on the aligned temperature sequence and the predicted temperature response sequence to generate a temperature residual sequence, and calculate the variance value of the temperature residual sequence by a sliding window to generate a residual fluctuation variance sequence. The temperature values ​​of the aligned temperature sequence and the predicted temperature response sequence are read according to the sampling point sequence number, and the point-by-point difference operation is performed to obtain the residual values ​​of the temperature residual sequence. Continuous sampling points are truncated along the time axis of the temperature residual sequence according to the window length to form segments, and the starting point of each segment is shifted according to the window step size. For any segment, the arithmetic mean of each residual value within the segment is calculated first, and then the difference between each residual value and the arithmetic mean is squared and the arithmetic mean is calculated again to obtain the variance value corresponding to that segment. The variance values ​​of each segment are arranged in segment order to generate the residual fluctuation variance sequence.

[0058] For example, the window length is continuous. Each sampling point and window step size is taken as continuous. When sampling the nth point, the temperature residual sequence is truncated sequentially. To the Sampling points are used as the first segment, the second segment, and the third segment. To the The sampling points are used as the second segment, and the variance values ​​of each segment are calculated and then arranged in the segment order to obtain the residual fluctuation variance sequence.

[0059] Using the variance values ​​of the residual fluctuation variance sequence as input, a kernel density estimation calculation is performed using a preset Gaussian kernel function. A weighted sum of the kernel function values ​​is calculated for each variance value point to generate a probability density distribution function. Variance values ​​are extracted from the residual fluctuation variance sequence, and a sample set of variance values ​​is constructed. Let the number of samples in the sample set be denoted as . And calculate the sample standard deviation of the variance set; then use the sample standard deviation and the sample size as a basis for calculation. Determine the bandwidth parameter and generate a sequence of variance points at fixed intervals within the variance range. For any variance point, calculate the difference between that variance point and each sample value in the variance sample set, divide by the bandwidth parameter to obtain a normalized difference, substitute the normalized difference into the Gaussian kernel function, and calculate the kernel function value. Summate all kernel function values ​​and perform normalization to obtain the density estimate for that variance point. Calculate the repeated density estimates for all variance points and arrange them in order of variance points to obtain the probability density distribution function. The Gaussian kernel density estimate is shown in the following formula: , in, This is the density estimate of the probability density distribution function at the points where the variance is taken; These are the variance points, generated at fixed intervals from the range of variance values. The number of samples is obtained by directly counting the number of samples in the variance sample set. The bandwidth parameter is determined by the variance, which is the sample standard deviation and the number of samples in the sample set. Together they determine and are used to control the smoothing scale; For the first Each variance sample value is composed of variance values ​​calculated piecewise from the residual fluctuation variance sequence; Pi is a constant. This is an exponential operation. The physical meaning is: each variance sample value is considered as a Gaussian shape density contribution centered on that sample value, and calculated according to the bandwidth parameter. Scaling the lateral extent of the density contribution, and then... The density contributions are summed and normalized to obtain a smooth probability density distribution of the residual fluctuation variance values ​​on a continuous number axis. Bandwidth parameter. The principle for selecting the value is: when the sample standard deviation is large or the sample size is small, take the larger value. To reduce density curve jitter; when the sample standard deviation is small or the sample size is large, take the smaller value. To preserve density details.

[0060] For example, extract from the residual fluctuation variance sequence The variance sample values ​​constitute a variance value sample set, and the sample standard deviation is calculated; within the variance value range, every [number] variance sample values ​​[are considered]. Generate a variance point For any variance value point Calculate separately and The normalized difference is obtained by dividing the difference of each variance sample value by the bandwidth parameter, substituting it into the exponent term, and summing the results to obtain the density estimate for that variance point. The probability density distribution function is obtained by repeatedly calculating and arranging all variance values.

[0061] The probability density distribution function is integrated to obtain the cumulative distribution function, and the lower and upper quantiles are solved on the cumulative distribution function to generate a dynamic confidence interval sequence.

[0062] The variance point sequence is sorted in ascending order, and the interval between adjacent variance points is used as the integration step size. For each variance point, the density estimate is sequentially multiplied by the integration step size to calculate the area increment, and the area increments are accumulated to obtain the cumulative distribution function. Lower and upper quantile probabilities are set, and the cumulative distribution function is iterated through. The variance point where the cumulative distribution function first exceeds or equals the lower quantile probability is identified as the lower quantile variance value, and the variance point where the cumulative distribution function first exceeds or equals the upper quantile probability is identified as the upper quantile variance value. The lower and upper quantile variance values ​​are output as interval endpoint pairs for the dynamic confidence interval sequence.

[0063] For example, the confidence level is set to Based on this, the quantile probability value is taken. The probability value of the upper quantile is Locking the corresponding values ​​on the cumulative distribution function respectively and The variance values ​​are taken as the lower quantile variance value and the upper quantile variance value, and the two are used to form the interval endpoint pair to output the dynamic confidence interval sequence.

[0064] Optionally, the generation probability density distribution function includes: Variance values ​​are extracted using the residual fluctuation variance sequence, a variance value sample set is constructed, and the sample standard deviation of the variance value sample set is calculated. Traverse the residual fluctuation variance sequence and read the variance values ​​one by one. Write all the variance values ​​into the variance value sample set in the order of reading. Calculate the arithmetic mean of the variance value sample set. Square the difference between each variance value and the arithmetic mean and sum the results. Divide the sum by the sample size minus one and perform a square root operation to obtain the sample standard deviation. The sample size is obtained by counting the number of samples in the variance value sample set.

[0065] For example, the residual fluctuation variance sequence contains When the variance value is 1, Each variance value is written into the variance value sample set in sequence, and the sample standard deviation is calculated by taking the square root of the variance value sample set by "sum of squared differences ÷ (Q-1)".

[0066] The bandwidth parameter is calculated based on the standard deviation of the sample, and the normalized difference between the variance value point and the sample value in the variance value sample set is calculated. The difference is then substituted into the Gaussian kernel function to obtain the kernel function value sequence. Read the sample standard deviation and the sample size, and then... Calculate the bandwidth parameters, where For sample standard deviation, Given the number of samples; generate a sequence of variance points at fixed intervals within the variance range; for any variance point, iterate through each variance sample value in the variance sample set, calculate the difference between the variance point and the variance sample value, divide by the bandwidth parameter to obtain the normalized difference, substitute the normalized difference into the Gaussian kernel function to obtain the kernel function value, and write all kernel function values ​​into the kernel function value sequence in the traversal order.

[0067] For example, when the sample standard deviation for And sample size for At that time, according to Calculate the bandwidth parameter; when the variance takes the value of... And the variance sample value is When calculating the normalized difference Substitute the values ​​into the Gaussian kernel function to obtain the corresponding kernel function values, and then iterate through all variance sample values ​​to obtain the kernel function value sequence.

[0068] The summation value is obtained by summing the kernel function value sequence, the number of samples is obtained by using the variance value sample set, the summation value is divided by the product of the number of samples and the bandwidth parameter to obtain the density estimate of the variance value points, and the probability density distribution function is constructed.

[0069] Summation is performed on all kernel function values ​​in the kernel function value sequence to obtain the summation value, and the sample size and bandwidth parameter of the variance value sample set are read. The summation value is divided by the product of the sample size and the bandwidth parameter to obtain the density estimate value corresponding to the variance value point. The kernel function value sequence and density estimate value calculation are repeated for each variance value point in the variance value point sequence, and the density estimates corresponding to each variance value point are arranged in the order of the variance value points to construct the probability density distribution function.

[0070] For example, when the variance value point sequence contains When the variance takes a value point, the variance is then taken into account sequentially for the first variance point. To the Calculate the density estimate for each variance point, and then... The probability density distribution function is obtained by arranging the density estimates in order of their variance values.

[0071] Optionally, the step of performing gradient descent updates on the multi-layer one-dimensional convolutional kernel using the topological violation feature vector includes: The residual fluctuation variance sequence and the dynamic confidence interval sequence are used to perform interval comparison, and the outbound index is marked to calculate the outbound point count, the maximum length of the continuous outbound segment and the maximum outbound amplitude, and to construct the topological violation feature vector. Sequence numbered by time segment Read the variance values ​​of the residual fluctuation variance sequence, and read the time segment numbers in the dynamic confidence interval sequence. The corresponding lower quantile variance and upper quantile variance are compared; an interval comparison is performed on the variance values. If the variance value is less than the lower quantile variance or greater than the upper quantile variance, the time segment number is changed. Write the data into the superboundary index set. Traverse all time segment indices and count the number of elements in the superboundary index set to obtain the superboundary point count. Sort the time segment indices in the superboundary index set from smallest to largest, and perform a continuity check ("whether it is equal to 1") on the difference between adjacent superboundary indices to obtain continuous superboundary segments. Count the length of each continuous superboundary segment and lock the maximum length to obtain the maximum length of the continuous superboundary segment. For the variance value corresponding to each superboundary index, calculate the absolute value of the difference between it and the variance value of the lower quantile point and the absolute value of the difference between it and the variance value of the upper quantile point. Take the larger of the two values ​​as the superboundary amplitude corresponding to the superboundary index, and lock the maximum value among all superboundary amplitudes as the maximum superboundary amplitude. Concatenate the superboundary point count, the maximum length of the continuous superboundary segment, and the maximum superboundary amplitude in a fixed order to construct the topological violation feature vector.

[0072] For example, when a certain time segment number The corresponding variance value is greater than the time segment number. When the corresponding upper quantile variance value is obtained, the time segment number is used. Write to the set of superbounded indices; after traversal, use the number of elements in the set of superbounded indices as the superbound point count, and use the difference between adjacent superbounded indices as the superbound point count. The maximum length of the continuous segment is taken as the maximum length of the continuous superboundary segment, and the maximum value of the superboundary amplitude is taken as the maximum value of the superboundary amplitude. The three are concatenated in order to obtain the topological violation feature vector.

[0073] Read the cluster center vectors from the historical fault feature library, calculate the Euclidean distance between the topological violation feature vector and the cluster center vector, perform numerical multiplication between the adaptive update step size coefficient and the weight gradient matrix, and perform parameter update for the multilayer one-dimensional convolution kernel.

[0074] The system sequentially reads cluster center vectors from the historical fault feature database and performs dimension-wise difference, squaring, summation, and square root operations on each cluster center vector and the topological violation feature vector to obtain the Euclidean distance. The cluster center vector corresponding to the minimum Euclidean distance is then locked as the locked cluster center vector. The adaptive update step size coefficients and weight gradient matrix are read, and each element of the weight gradient matrix is ​​multiplied by the adaptive update step size coefficient to generate a weighted gradient matrix. Update operations are performed on the weight values ​​of the multi-layer one-dimensional convolution kernel. The update operation involves subtracting the corresponding element's value from the current weight value to obtain the updated weight value, which is then written into the convolution kernel weight matrix to complete the parameter update. By calculating the above Euclidean distance and feature vectors, fault features can be located and parameter corrections can be triggered, such as... Figure 3 As shown, the upper subplot illustrates the adaptive update step size coefficient that changes with the time window, reflecting the model's sensitivity to outlier data. The lower subplot uses a bubble matrix to show the energy intensity of different modal components in each time window, i.e., bubble size, and the degree of topological violation, i.e., grayscale depth, presenting the time-frequency distribution of fault characteristics and their driving effect on model evolution.

[0075] For example, when an element in the weight gradient matrix is And the adaptive update step size coefficient is At that time, and Multiplying yields the weighted gradient value. When the corresponding one-dimensional convolution kernel weight value is When, execute The updated weight values ​​are obtained, and the update operation is repeated for all weight values ​​to obtain the updated convolution kernel weight matrix.

[0076] Based on the same inventive concept, this invention also provides a powder equipment vibration-temperature fusion fault diagnosis system, such as... Figure 4 As shown, the system includes: The vibration and temperature sequence acquisition and alignment module is used to acquire vibration acceleration sequences and surface temperature sequences of the operating parts of the powder equipment, perform cross-correlation calculations on the vibration acceleration sequences and the surface temperature sequences, record the time delay of the peak values, and perform resampling and translation operations on the vibration acceleration sequences based on the time delay to generate aligned vibration sequences and aligned temperature sequences. The vibration energy flow construction module is used to perform variational mode decomposition and sliding window root mean square calculation on the aligned vibration sequence to obtain a multi-band vibration intensity matrix, and to perform weighted summation on the multi-band vibration intensity matrix through the modal center frequencies to construct a vibration energy flow sequence. The thermal response convolution prediction module is used to construct a sliding window sample set using the vibration energy flow sequence and the aligned temperature sequence, initialize a multi-layer one-dimensional convolution kernel, perform error backpropagation and parameter iterative update on the sliding window sample set, generate a convolution kernel weight matrix, and perform convolution operation on the vibration energy flow sequence using the convolution kernel weight matrix to generate a predicted temperature response sequence. The dynamic confidence boundary generation module is used to perform point-by-point difference and variance statistical operations on the aligned temperature sequence and the predicted temperature response sequence to generate a residual fluctuation variance sequence, construct a probability density distribution function on the residual fluctuation variance sequence based on the kernel density estimation method, and perform integration and quantile solving on the probability density distribution function to generate a dynamic confidence interval sequence. The fault identification and parameter update module is used to extract the topological violation feature vectors of the residual fluctuation variance sequence whose amplitudes exceed the dynamic confidence interval sequence, calculate the Euclidean distance between the topological violation feature vectors and each cluster center in the preset historical fault feature library, and use the topological violation feature vectors to perform gradient descent update on the multilayer one-dimensional convolution kernel.

[0077] It should be noted that the functional division and information interaction between the various modules described above are logical, but in terms of physical implementation, they can be integrated on the same software platform or deployed in a distributed manner. The connections between them represent data flow and control flow, aiming to collaboratively achieve the objectives of this invention. The above descriptions are merely exemplary embodiments of this invention and should not be construed as limiting the scope of protection of this invention.

Claims

1. A method for diagnosing faults in the vibration-temperature fusion process of powder equipment, characterized in that, The method includes: Vibration acceleration sequences and surface temperature sequences are collected from the operating components of the powder equipment. Cross-correlation calculations are performed on the vibration acceleration sequences and surface temperature sequences, and the time delay of the peak values ​​is recorded. Based on the time delay, resampling and translation operations are performed on the vibration acceleration sequences to generate aligned vibration sequences and aligned temperature sequences. Variational mode decomposition and sliding window root mean square calculation are performed on the aligned vibration sequence to obtain a multi-band vibration intensity matrix. The multi-band vibration intensity matrix is ​​then weighted and summed using the modal center frequencies to construct a vibration energy flow sequence. A sliding window sample set is constructed using the vibration energy flow sequence and the aligned temperature sequence. Multi-layer one-dimensional convolution kernels are initialized to perform error backpropagation and parameter iterative update on the sliding window sample set, generating a convolution kernel weight matrix. The convolution operation is then performed on the vibration energy flow sequence using the convolution kernel weight matrix to generate a predicted temperature response sequence. Perform point-by-point difference and variance statistics on the aligned temperature sequence and the predicted temperature response sequence to generate a residual fluctuation variance sequence. Construct a probability density distribution function on the residual fluctuation variance sequence based on the kernel density estimation method, and perform integration and quantile solving on the probability density distribution function to generate a dynamic confidence interval sequence. Extract the topological violation feature vectors from the residual fluctuation variance sequence whose amplitudes exceed the dynamic confidence interval sequence, calculate the Euclidean distance between the topological violation feature vectors and each cluster center in the preset historical fault feature library, and use the topological violation feature vectors to perform gradient descent update on the multilayer one-dimensional convolution kernel.

2. The method for diagnosing vibration-temperature fusion faults in powder equipment according to claim 1, characterized in that, The generation of the aligned vibration sequence and the aligned temperature sequence includes: The vibration acceleration sequence and the surface temperature sequence are subjected to mean reduction and fast Fourier transform to obtain the vibration spectrum sequence and the temperature spectrum sequence. The conjugate product is calculated using the vibration spectrum sequence and the temperature spectrum sequence, and an inverse fast Fourier transform is performed on the conjugate product to generate a sequence of cross-correlation functions. Traverse the amplitude of the cross-correlation function sequence, extract the x-coordinate index value of the point with the maximum amplitude, convert the x-coordinate index value into a time unit, and obtain the time delay. The vibration acceleration sequence is resampled by linear interpolation using the time delay to generate an intermediate sequence. The intermediate sequence is then subjected to a reverse time displacement operation to generate an aligned vibration sequence that is phase-synchronized with the surface temperature sequence. The aligned temperature sequence is then obtained using the surface temperature sequence.

3. The method for diagnosing vibration-temperature fusion faults in powder equipment according to claim 1, characterized in that, The constructed vibration energy flow sequence includes: A variational constraint optimization equation is constructed for the aligned vibration sequence. The variational constraint optimization equation is then subjected to frequency domain iterative search using a preset alternating direction multiplier algorithm to decompose and generate modal components. The modal components are sliced ​​using a time-domain sliding window, the root mean square amplitude of each slice is calculated, and the slices are arranged according to the modal number and time dimension to generate a multi-band vibration intensity matrix. For each modal component, the center frequency value is extracted, a frequency weighting vector is constructed, and matrix multiplication is performed between the multi-band vibration intensity matrix and the frequency weighting vector to generate a vibration energy flow sequence.

4. The method for diagnosing vibration-temperature fusion faults in powder equipment according to claim 3, characterized in that, The decomposition generates modal components including: Perform a fast Fourier transform on the aligned vibration sequence to generate a vibration frequency domain sequence; By introducing Lagrange multiplier variables, an augmented Lagrange function containing modal spectrum variables and modal center frequency variables is constructed using the vibration frequency domain sequence; The augmented Lagrange function is subjected to the alternating direction multiplier algorithm iterative operation. In each iteration, the modal spectrum variable, the modal center frequency variable, and the Lagrange multiplier variable are updated sequentially until the convergence condition is met, and the converged modal spectrum sequence is output. Perform an inverse fast Fourier transform on the convergent mode spectrum sequence to generate mode components.

5. The method for diagnosing vibration-temperature fusion faults in powder equipment according to claim 2, characterized in that, The generation of the predicted temperature response sequence includes: Forward propagation calculations are performed on the vibration energy flow sequence fragments and aligned temperature sequence fragments in the sliding window sample set. Convolution operations are performed on the vibration energy flow sequence fragments through the multi-layer one-dimensional convolution kernel to generate intermediate predicted temperature values. Calculate the mean squared error loss function between the predicted temperature median and the aligned temperature sequence segment, and perform partial derivative calculation of the mean squared error loss function with respect to the multilayer one-dimensional convolution kernel based on the chain rule to generate a weight gradient matrix; The weight gradient matrix is ​​used to perform gradient descent update on the multi-layer one-dimensional convolution kernel to generate a convolution kernel weight matrix. The convolution kernel weight matrix is ​​then used to perform point-by-point sliding convolution on the vibration energy flow sequence to generate a predicted temperature response sequence.

6. The method for diagnosing vibration-temperature fusion faults in powder equipment according to claim 5, characterized in that, The method further includes: The maximum amplitude value is extracted from the cross-correlation function sequence as a correlation strength scalar, and the numerical result of the mean squared error loss function is used as a prediction error scalar. The correlation strength scalar and the prediction error scalar are used to perform a ratio calculation to generate an adaptive update step size coefficient.

7. The method for diagnosing vibration-temperature fusion faults in powder equipment according to claim 1, characterized in that, The generation of the dynamic confidence interval sequence includes: Perform point-by-point difference operation on the aligned temperature sequence and the predicted temperature response sequence to generate a temperature residual sequence, and calculate the variance value of the temperature residual sequence by a sliding window to generate a residual fluctuation variance sequence. Using the variance values ​​of the residual fluctuation variance sequence as input, a kernel density estimation calculation is performed using a preset Gaussian kernel function. A weighted sum of the kernel function values ​​is calculated for each variance value point to generate a probability density distribution function. The probability density distribution function is integrated to obtain the cumulative distribution function, and the lower and upper quantiles are solved on the cumulative distribution function to generate a dynamic confidence interval sequence.

8. The method for diagnosing vibration-temperature fusion faults in powder equipment according to claim 7, characterized in that, The generation probability density distribution function includes: Variance values ​​are extracted using the residual fluctuation variance sequence, a variance value sample set is constructed, and the sample standard deviation of the variance value sample set is calculated. The bandwidth parameter is calculated based on the standard deviation of the sample, and the normalized difference between the variance value point and the sample value in the variance value sample set is calculated. The difference is then substituted into the Gaussian kernel function to obtain the kernel function value sequence. The summation value is obtained by summing the kernel function value sequence, the number of samples is obtained by using the variance value sample set, the summation value is divided by the product of the number of samples and the bandwidth parameter to obtain the density estimate of the variance value points, and the probability density distribution function is constructed.

9. A method for diagnosing vibration-temperature fusion faults in powder equipment according to claim 6, characterized in that, The step of updating the gradient descent of the multi-layer one-dimensional convolutional kernel using the topological violation feature vector includes: The residual fluctuation variance sequence and the dynamic confidence interval sequence are used to perform interval comparison, and the outbound index is marked to calculate the outbound point count, the maximum length of the continuous outbound segment and the maximum outbound amplitude, and to construct the topological violation feature vector. Read the cluster center vectors from the historical fault feature library, calculate the Euclidean distance between the topological violation feature vector and the cluster center vector, perform numerical multiplication between the adaptive update step size coefficient and the weight gradient matrix, and perform parameter update for the multilayer one-dimensional convolution kernel.

10. A powder equipment vibration-temperature fusion fault diagnosis system, applied to the powder equipment vibration-temperature fusion fault diagnosis method as described in any one of claims 1-9, characterized in that, The system includes: The vibration and temperature sequence acquisition and alignment module is used to acquire vibration acceleration sequences and surface temperature sequences of the operating parts of the powder equipment, perform cross-correlation calculations on the vibration acceleration sequences and the surface temperature sequences, record the time delay of the peak values, and perform resampling and translation operations on the vibration acceleration sequences based on the time delay to generate aligned vibration sequences and aligned temperature sequences. The vibration energy flow construction module is used to perform variational mode decomposition and sliding window root mean square calculation on the aligned vibration sequence to obtain a multi-band vibration intensity matrix, and to perform weighted summation on the multi-band vibration intensity matrix through the modal center frequencies to construct a vibration energy flow sequence. The thermal response convolution prediction module is used to construct a sliding window sample set using the vibration energy flow sequence and the aligned temperature sequence, initialize a multi-layer one-dimensional convolution kernel, perform error backpropagation and parameter iterative update on the sliding window sample set, generate a convolution kernel weight matrix, and perform convolution operation on the vibration energy flow sequence using the convolution kernel weight matrix to generate a predicted temperature response sequence. The dynamic confidence boundary generation module is used to perform point-by-point difference and variance statistical operations on the aligned temperature sequence and the predicted temperature response sequence to generate a residual fluctuation variance sequence, construct a probability density distribution function on the residual fluctuation variance sequence based on the kernel density estimation method, and perform integration and quantile solving on the probability density distribution function to generate a dynamic confidence interval sequence. The fault identification and parameter update module is used to extract the topological violation feature vectors of the residual fluctuation variance sequence whose amplitudes exceed the dynamic confidence interval sequence, calculate the Euclidean distance between the topological violation feature vectors and each cluster center in the preset historical fault feature library, and use the topological violation feature vectors to perform gradient descent update on the multilayer one-dimensional convolution kernel.

Images