Powder packaging machine transmission component fault pre-diagnosis method based on vibration signal analysis

Abstract

The present application relates to a kind of powder packaging machine transmission component fault pre-diagnosis method based on vibration signal analysis, the present application is accurate to collect multiple source vibration signals, standardization processing of time-frequency characteristics, generate fault semantic anchor point library in combination with prior knowledge of dynamics and under the training method of multiple loss constraint convolutional neural network, realize the adaptive alignment of vibration signal hidden space feature and physical fault mechanism.The present application outputs the joint attribution time-frequency graph of specific fault type, and the coincidence degree of model attention area and theoretical fault frequency label is based on physical consistency check, significantly improve diagnostic accuracy and explainability, effectively assist equipment operation decision, the present application is by constructing the fault semantic anchor point library based on physical mechanism, and it is embedded in the hidden space optimization process of convolutional neural network in differentiable way, effectively overcome the problem that traditional data-driven model is generally present in powder packaging machine transmission chain fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis and vibration signal interpretability analysis technology for intelligent manufacturing equipment, and in particular to a method for fault prediction of transmission components of a powder packaging machine based on vibration signal analysis. Background Technology

[0002] Currently, fault diagnosis of transmission components in intelligent manufacturing equipment, especially complex mechanical systems such as powder packaging machines, primarily relies on vibration signal analysis as its core technical approach. Traditional solutions commonly employ methods such as signal enhancement, wavelet packet decomposition, empirical mode decomposition (EMD), time-frequency domain energy clustering, envelope spectrum analysis, and Hilbert-Huang transform, combined with classification techniques like support vector machines (SVM), decision trees, and shallow neural networks, to achieve anomaly detection and fault type identification at the component level and component hierarchy. In recent years, deep learning models such as convolutional neural networks (CNN), recurrent neural networks (RNN), and self-attention Transformer structures have become the mainstream trend for intelligent identification of subtle faults in industrial equipment due to their significantly improved ability to automatically mine features from high-dimensional, complex, and nonlinear signals. In industrial settings, an increasing number of transmission chain monitoring systems have incorporated end-to-end deep fault diagnosis modules, enabling data-driven decision-making across the entire process using multi-source, multi-point vibration data.

[0003] However, existing technologies have the following core problems: First, while deep learning models significantly improve the sensitivity of detecting subtle early-stage faults in transmission components of powder packaging machines, such as geared motors, synchronous belts, gearboxes, and bearings, their judgment process exhibits typical "black box" characteristics. Models often only output final classification labels, lacking explanations of the physical traceability underlying the decision-making process. Engineers often struggle to understand why the model classifies a particular signal as corresponding to a specific component's fault type, and cannot determine whether the signal region the model focuses on aligns with actual mechanical mechanisms (such as theoretical characteristic frequencies, modulation sidebands, and harmonic positions).

[0004] Second, to address complex background noise and weak feature perturbations, some literature introduces multi-level preprocessing such as signal enhancement, mode decomposition, feature reconstruction, and transfer fine-tuning. However, these approaches not only increase implementation complexity but may also introduce secondary information distortion, making the physical meaning of the final diagnostic output increasingly ambiguous and weakening the trust of the on-site maintenance team. Especially for early, minor faults, the high-dimensional abstract features of deep models have an increasingly severe "shielding effect" on the physical mechanism, ultimately resulting in diagnostic results that cannot provide effective support for maintenance decisions.

[0005] Third, while some domestic and international patents and papers have attempted to introduce feature importance assessment (such as Grad-CAM and visual activation hot zones) into the vibration signal recognition process, these methods mostly remain at the level of highlighting signals in a statistical sense. They fail to effectively link physical constraints such as the dynamic characteristic frequencies, theoretical stress responses, and typical fault mechanisms of various transmission components with the deep network reasoning process. Interpretable solutions lacking physical prior constraints struggle to meet industrial users' demands for traceability and auditability of diagnostic outputs.

[0006] Fourth, with the increasing demand for intelligent equipment and predictive maintenance, field customers are placing higher demands on the "trust" of AI fault diagnosis tools and the boundaries of responsibility for misjudgments and omissions. Existing data-driven technologies have failed to couple the interpretability of vibration signals with physical consistency, resulting in insufficient transparency and execution efficiency in the maintenance decision-making chain, which is not conducive to the large-scale and efficient promotion of complex transmission chains. Summary of the Invention

[0007] This invention provides a method for predicting faults in the transmission components of a powder packaging machine based on vibration signal analysis, aiming to solve the problems existing in the prior art mentioned in the background section.

[0008] The technical solution of the present invention is: a method for pre-diagnosing faults of transmission components of a powder packaging machine based on vibration signal analysis, S1: collecting the vibration time-domain signal of the transmission components of the powder packaging machine under rated working conditions, and obtaining the original vibration signal data; S2: Perform bandpass filtering and normalization preprocessing on the original vibration signal data to eliminate background noise interference and generate a preprocessed vibration signal; S3: Based on the multi-level dynamic constraints of the transmission components of the powder packaging machine, the theoretical fault feature frequency families of the geared motor, synchronous belt, gearbox, bearing and cam mechanism are extracted, and each frequency family is discretized into a sparse vector to construct a fault semantic anchor point library. S4: Input the preprocessed vibration signal into the convolutional neural network model. During the training phase, calculate the KL divergence loss between the feature vector and the fault semantic anchor library through the frequency domain projection head, and optimize the parameters of the convolutional neural network model in conjunction with the cross-entropy loss to generate a semantic anchor convolutional neural network model. S5: In the fault diagnosis reasoning stage, the preprocessed vibration signal is input into the semantic anchored convolutional neural network model to obtain the penultimate layer feature map; S6: Based on the penultimate layer feature map and the final classification layer output, the gradient weighted class activation mapping algorithm is used to calculate the gradient heatmap; S7: The gradient heatmap is inversely mapped to the time-frequency plane via short-time Fourier transform, and the physical frequency location labels corresponding to the fault type in the fault semantic anchor point library are superimposed to generate a joint attribution map. S8: Determine whether the overlap between the model's region of interest and the physical region in the joint attribution graph meets the preset threshold. If not, output a physical inconsistency alarm.

[0009] Furthermore, in step S1, the triaxial accelerometer is driven by the acquisition control command to capture the analog mechanical vibration of the transmission component at the rated speed, and the analog quantity is quantized into a discrete time series through the analog-to-digital conversion module to generate a single-channel raw vibration digital signal; timestamp verification and packet loss detection are performed on each single-channel raw vibration digital signal to remove abnormal data segments caused by transmission jitter, and the time reference of the multi-channel data is aligned to generate a spatiotemporally aligned multi-source vibration signal set.

[0010] Furthermore, in step S2, the time-domain discrete sequence of the multi-source vibration signal set is converted into a frequency-domain power spectral density distribution using the fast Fourier transform algorithm to generate an initial frequency-domain feature vector. Based on the non-fault background noise frequency band boundary of the initial frequency domain feature vector, a Butterworth digital bandpass filter transfer function with steep cutoff characteristics is constructed, and a convolution filtering operation is performed on the original vibration signal data to generate a bandpass filtered vibration signal. For the time-domain waveform envelope of the bandpass filtered vibration signal, the Hilbert transform algorithm is used to extract the instantaneous amplitude sequence, and the global root mean square value is calculated as a dynamic scaling reference to generate a normalized reference scalar. The normalized reference scalar is used to perform element-wise division on each sampling point of the bandpass filtered vibration signal, and the result is mapped to the standard normal distribution interval.

[0011] Furthermore, in step S6, the penultimate layer feature map is forward-propagated to obtain the category probability distribution vector of the final classification layer; Based on the scalar output value of the target fault category in the category probability distribution vector, perform backpropagation operation to calculate the gradient tensor of the target fault category relative to each spatial location pixel in the penultimate feature map; Perform a global average pooling operation on the gradient tensor along the channel dimension to generate a channel importance weight vector corresponding to the target fault category; The channel importance weight vector and the penultimate layer feature map are fused element-wise, and nonlinear activation is performed through a linear rectified function to generate a preliminary fault attention response heatmap. Bilinear interpolation upsampling is performed on the preliminary fault attention response heatmap to generate a spatially aligned gradient heatmap.

[0012] The beneficial technical effects of this invention are as follows: 1) This invention effectively overcomes the technical defects of poor interpretability and deviation of diagnostic logic from mechanical principles that are common in traditional data-driven models for fault diagnosis of transmission chains in powder packaging machines by constructing a fault semantic anchor point library based on physical mechanisms and embedding it into the latent space optimization process of convolutional neural networks in a differentiable manner. 2) This invention discretizes the theoretical fault frequency family of key components such as geared motors, synchronous belts, gearboxes, bearings, and cam mechanisms into sparse vectors, and uses them as supervisory signals to guide the learning direction of the last layer of feature representation in CNN. This allows the model to explicitly encode physically meaningful frequency domain preferences during the classification decision process. Combined with a weighted KL divergence loss and cross-entropy joint optimization strategy, this invention significantly improves the model's sensitivity and focusing ability to key frequency bands without introducing additional labeling costs. This ensures that the fault representations it learns not only conform to the data distribution pattern but also remain consistent with the dynamic characteristics of the equipment, thereby greatly enhancing the credibility and engineering applicability of the diagnostic results. 3) This invention designs a dual-path attribution fusion mechanism to achieve collaborative visualization of data-driven attention and physical prior knowledge in the time-frequency domain. During the inference stage, this invention simultaneously activates the data-driven sensitive region generated by Grad-CAM++ and the physical anchor mask constructed based on the theoretical fault frequency, and performs spatiotemporal alignment fusion through a dynamic weight adjustment strategy to output a joint attribution graph that combines model attention and physical requirements. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the method of an embodiment of the present invention; Figure 2 A schematic diagram illustrating the process of constructing a fault semantic anchor point library for an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the process of generating a semantically anchored convolutional neural network model for an embodiment of the present invention. Detailed Implementation

[0014] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0015] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0016] like Figure 1 As shown in the figure, this embodiment provides a method for fault prediction of transmission components of a powder packaging machine based on vibration signal analysis, which specifically includes the following steps: S1: Collect the vibration time-domain signal of the transmission components of the powder packaging machine under rated operating conditions and obtain the raw vibration signal data; S2: Perform bandpass filtering and normalization preprocessing on the original vibration signal data to eliminate background noise interference and generate a preprocessed vibration signal; S3: Based on the multi-level dynamic constraints of the transmission components of the powder packaging machine, the theoretical fault feature frequency families of the geared motor, synchronous belt, gearbox, bearing and cam mechanism are extracted, and each frequency family is discretized into sparse vectors to construct a fault semantic anchor point library; S4: Input the preprocessed vibration signal into the convolutional neural network model. During the training phase, calculate the KL divergence loss between the feature vector and the fault semantic anchor library through the frequency domain projection head, and optimize the model parameters in conjunction with the cross-entropy loss to generate a semantic anchor convolutional neural network model. S5: In the fault diagnosis reasoning stage, the preprocessed vibration signal is input into the semantic anchored convolutional neural network model to obtain the penultimate layer feature map; S6: Based on the penultimate layer feature map and the final classification layer output, the gradient weighted class activation mapping algorithm is used to calculate the gradient heatmap; S7: The gradient heatmap is inversely mapped to the time-frequency plane via short-time Fourier transform, and the physical frequency location labels corresponding to the fault type in the fault semantic anchor point library are superimposed to generate a joint attribution map. S8: Determine whether the overlap between the model's region of interest and the physical region in the joint attribution graph meets the preset threshold. If not, output a physical inconsistency alarm.

[0017] In this embodiment, step S1 involves acquiring the vibration time-domain signal of the transmission component of the powder packaging machine under rated operating conditions to obtain the original vibration signal data. This specifically includes the following steps: S1.1: Based on the transmission chain topology of the powder packaging machine, vibration sensitivity analysis is performed on key measuring points such as the output end of the geared motor, the synchronous belt tensioner, the gearbox bearing seat, and the main shaft of the filling turntable to determine the installation position and axial layout of the triaxial accelerometer and generate a sensor layout scheme. S1.2: Based on the sensor deployment scheme, the multi-channel data acquisition card is triggered and aligned using the industrial Ethernet clock synchronization protocol. The Nyquist sampling parameters with a sampling frequency ten times higher than the highest fault characteristic frequency are set to generate synchronous acquisition control commands. S1.3: Responding to the synchronous acquisition control command, the triaxial accelerometer is driven to capture the analog mechanical vibration of the transmission component at the rated speed, and the analog quantity is quantized into a discrete time series through the analog-to-digital conversion module to generate a single-channel original digital vibration signal; S1.4: Perform timestamp verification and packet loss detection on each single-channel raw vibration digital signal, remove abnormal data segments caused by transmission jitter, align the time base of multi-channel data, and generate a spatiotemporally aligned multi-source vibration signal set; S1.5: Based on a spatiotemporally aligned multi-source vibration signal set, a sliding window truncation technique is used to extract continuous data segments containing complete rotation cycles, and metadata tags are encapsulated to form standardized data objects, generating raw vibration signal data for subsequent preprocessing.

[0018] In this embodiment, step S2 involves bandpass filtering and normalization preprocessing of the original vibration signal data to eliminate background noise interference and generate a preprocessed vibration signal. This specifically includes the following steps: S2.1: Based on the spatiotemporally aligned multi-source vibration signal set generated in the previous steps, the time-domain discrete sequence is converted into a frequency-domain power spectral density distribution using the fast Fourier transform algorithm to identify the main energy concentration frequency band and non-fault background noise frequency band under the rated working conditions of the transmission components of the powder packaging machine, and generate an initial frequency-domain feature vector containing frequency amplitude information. S2.2: Based on the non-fault background noise frequency band boundary identified in the initial frequency domain feature vector, construct the transfer function of the Butterworth digital bandpass filter with steep cutoff characteristics, and perform convolution filtering operation on the original vibration signal data to remove low-frequency mechanical shaking interference and high-frequency electronic thermal noise, and generate a bandpass filtered vibration signal with significantly improved signal-to-noise ratio. S2.3: For the time-domain waveform envelope of the bandpass filtered vibration signal, the Hilbert transform algorithm is used to extract the instantaneous amplitude sequence and calculate the global root mean square value as a dynamic scaling reference to quantify the energy difference between different acquisition channels and generate a normalized reference scalar characterizing the overall energy level of the signal.

[0019] S2.4: Element-wise division is performed on each sampling point of the bandpass filtered vibration signal using a normalized reference scalar, and the result is mapped to the standard normal distribution interval with zero mean and unit variance to eliminate the inconsistency of dimensions caused by differences in sensor sensitivity and generate standardized vibration time series data with uniform statistical characteristics. S2.5: Based on standardized vibration time series data, a sliding window overlapping truncation process is performed to segment the continuous long sequence into fixed-length data frames that meet the input dimension requirements of the convolutional neural network, and timestamp metadata is encapsulated to form a standard data object, ultimately generating a preprocessed vibration signal for subsequent semantic anchoring model input.

[0020] In this embodiment, in step S3, based on the multi-level dynamic constraints of the transmission components of the powder packaging machine, the theoretical fault feature frequency families of the reduction motor, synchronous belt, gearbox, bearing, and cam mechanism are extracted, and each frequency family is discretized into a 128-dimensional sparse vector to construct a fault semantic anchor point library, such as... Figure 2 As shown, the specific steps include the following: S3.1: Based on the topological parameters and rated operating speed data of the transmission chain of the powder packaging machine, the unbalanced fundamental frequency and its harmonics of the geared motor rotor, the slippage impact frequency and its sideband of the synchronous belt, the meshing frequency of each stage of the gearbox and its modulation sideband, the fault characteristic frequencies of the inner and outer rings and rolling elements of the bearing, and the impact transient envelope frequency of the follower of the cam mechanism are analyzed and calculated to generate a theoretical fault characteristic frequency set covering the entire transmission chain. S3.2: Based on the generated set of theoretical fault characteristic frequencies, according to the ISO 10816 vibration standard and combined with the actual geometric parameters of the equipment, the tolerance range of the characteristic frequency values ​​corresponding to each type of fault is defined and the bandwidth is expanded to eliminate the frequency drift caused by manufacturing tolerances and operational fluctuations, and generate a list of fault characteristic frequency band intervals with engineering robustness. Furthermore, the frequency offset tolerance is numerically solved using a tolerance calculation method based on manufacturing tolerances and operating condition fluctuations, and the tolerance amplitude for each frequency is obtained. The following tolerance calculation formula is used: ; in, These are the theoretical characteristic frequency values; This refers to the speed fluctuation ratio; This refers to the number of teeth on the gear. For gear tooth tolerance; The diameter of the bearing rolling element; For the diameter tolerance of the bearing rolling elements; For the synchronous belt length tolerance; This refers to the length of the synchronous belt.

[0021] S3.3: Using a preset 128-dimensional zero-initialization vector as the basic mapping space, the center frequency of each frequency band in the generated fault feature frequency band interval list is mapped to the index position corresponding to the vector, and a unit weight value is assigned to the mapping position while the other positions remain zero. Sparse coding operation is performed to convert continuous frequency domain information into discrete digital expression, and the original sparse feature vector corresponding to a single type of fault is generated. S3.4: Based on the labels of various typical fault modes that may occur in powder packaging machines, the generated original sparse feature vectors of various single-type faults are subjected to category index association and standardized encapsulation processing. The sparse feature vectors of all fault types are arranged in order of fault type number and integrated into a unified matrix structure to generate a multi-dimensional fault semantic anchor matrix data object. S3.5: Perform persistent storage and metadata annotation operations on the generated multidimensional fault semantic anchor point matrix data object, add structured label information including transmission component names, fault mechanism descriptions and frequency calculation basis, form a standard fault semantic anchor point library that can be directly called by the frequency domain projection head of the convolutional neural network, and complete the transformation of physical prior knowledge into model-recognizable digital assets.

[0022] In this embodiment, in step S4, the preprocessed vibration signal is input into the convolutional neural network model. During the training phase, the KL divergence loss between the feature vector and the fault semantic anchor library is calculated using a frequency domain projection head, and the model parameters are optimized by combining the cross-entropy loss to generate a semantically anchored convolutional neural network model, such as... Figure 3 As shown, the specific steps include the following: S4.1: Based on the fault semantic anchor point library and preprocessed vibration signal generated in the previous steps, a convolutional neural network architecture containing a learnable frequency domain projection head is constructed. The frequency domain projection head is configured to receive the feature vector after global average pooling and output a sparse probability distribution vector with the same dimension as the fault semantic anchor point library, so as to establish a mapping channel between latent space features and physical fault frequency family. Furthermore, by utilizing the fully connected mapping layer and nonlinear activation function within the frequency domain projection head, the mapping from the global average pooling vector to the sparse probability distribution vector is achieved, ensuring that the output vector completely corresponds to the fault semantic anchor point library in both the numerical domain and index space. The probability distribution can be calculated as follows: ; in, It is a sparse probability distribution vector. For the Sigmoid function, Here is the projection head weight matrix. For pooling vectors, This is the bias vector.

[0023] Furthermore, a vector dimension consistency check algorithm is used to match the frequency domain projection head output vector with the sparse vector of the target fault category in the fault semantic anchor point library to ensure that the operands of the subsequent KL divergence loss calculation are completely consistent in terms of dimension, numerical domain and index correspondence.

[0024] For example, in the early bearing fault detection scenario of a powder packaging machine, the preprocessed vibration signal is a standardized data frame with a sampling frequency of 25600Hz and a window length of 1024 points. The convolutional neural network backbone is loaded with three one-dimensional convolutional layers (with kernel sizes of 64, 32, and 16, and a stride of 2 for each layer). After global average pooling, a feature vector of length 128 is generated. The frequency domain projection head contains a set of weight matrices. The size is 128×128, and the bias vector is... The length is 128, and the output is a sparse probability distribution vector after applying Sigmoid activation. Most of its elements are close to zero, with only high response values ​​close to 1 output at the index positions corresponding to the characteristic frequencies of bearing failures. This... Compared with the sparse vectors of bearing bpfo frequency categories in the fault semantic anchor library, the KL divergence loss is significantly lower than the model without semantic anchoring, which verifies the effectiveness of the frequency domain projection head in mapping latent space features to physical frequency families, and improves the traceability of fault decision-making basis in subsequent training.

[0025] S4.2: Perform multi-layer convolution and pooling operations on the input preprocessed vibration signal to extract deep time-frequency feature maps, and perform global average pooling on the deep time-frequency feature maps to generate feature vectors to be projected, which serve as direct input data for the frequency domain projection head; S4.3: The frequency domain projection head is used to perform nonlinear activation and element-wise weighting on the feature vector to be projected, generating a model-predicted sparse probability distribution vector. The physical constraint divergence loss value is calculated based on the model-predicted sparse probability distribution vector and the fault semantic anchor vector corresponding to the current sample, so as to quantify the degree of deviation between the model attention and the mechanical fault mechanism. For the feature vector to be projected generated by global average pooling, the Sigmoid nonlinear activation function is used to compress and map the continuous real feature values ​​to the [0,1] probability domain.

[0026] Furthermore, physical prior preferences are applied to each frequency dimension of the feature vector through element-wise weighting, resulting in a weighted frequency domain response vector.

[0027] Furthermore, the Softmax normalization algorithm is used to achieve exponential scaling normalization of the probabilities of each frequency dimension, and a sparse probability distribution vector is generated. This is to ensure that the predicted vectors satisfy the probability distribution constraints.

[0028] Furthermore, the Kullback–Leibler divergence formula is used to analyze the sparse probability distribution vector. The fault semantic anchor vector corresponding to the current sample Perform physical constraint divergence calculation: ; in For the first The model predicts a sparse probability distribution vector for each sample. This is the one-hot anchor probability vector corresponding to the fault type.

[0029] S4.4: Calculate the cross-entropy classification loss value based on the output of the final classification layer and the real fault labels, and then weight and fuse the physical constraint divergence loss value with the cross-entropy classification loss value to generate the joint optimization total loss value, which serves as the error metric benchmark for the backpropagation algorithm; Furthermore, a weighted fusion method is employed to synthesize the physical constraint divergence loss value and the cross-entropy loss value into a joint optimization total loss value, and a single scalar is generated as the error metric for the backpropagation stage, namely: ; in, This represents the physical constraint divergence loss value. This represents the cross-entropy loss value. and Let be the weighting coefficients, and let their values ​​satisfy . .

[0030] S4.5: Based on the joint optimization total loss value, perform gradient backpropagation and parameter update operations, iteratively adjust the weights of the convolutional neural network and the parameters of the frequency domain projection head until convergence is achieved to generate a semantically anchored convolutional neural network model, ensuring that the model has both high-precision fault identification capability and a feature response mechanism that conforms to the dynamic principle. Based on the joint optimization total loss value as input, the backpropagation algorithm is used to calculate and update the gradients of the weights of each layer in the convolutional neural network. Furthermore, the chain rule is used to perform gradient transfer operations on the convolutional kernel weights, bias parameters, and sparse mapping matrices in the frequency domain projection head, respectively, and the gradient tensors of each parameter in the current iteration are obtained.

[0031] Furthermore, a weight update formula is used for parameter correction: ; in Weights to be updated For learning rate, This corresponds to the gradient tensor. The same update formula is applied to each sparse weight position of the frequency domain projector to ensure that the physical constraint loss term effectively corrects the feature response preference.

[0032] In this embodiment, in step S5, during the fault diagnosis reasoning stage, the preprocessed vibration signal is input into the semantic anchored convolutional neural network model to obtain the penultimate layer feature map, specifically including the following steps: S5.1: Perform tensor dimension reconstruction processing on the preprocessed vibration signal to convert the one-dimensional time domain data into a three-dimensional tensor format that meets the input requirements of the convolutional neural network, and generate a standardized input tensor. S5.2: Perform multi-layer convolution kernel sliding weighted summation operation based on standardized input tensor to extract local high-frequency impact features and low-frequency modulation modes in vibration signals and generate multi-channel primary feature mapping map; S5.3: Use the batch normalization algorithm to perform distribution shifting and scaling correction on the multi-channel primary feature map to eliminate the internal covariate offset caused by the difference in training data distribution and generate a normalized intermediate feature matrix; S5.4: Threshold truncation and sparsification are performed on the normalized intermediate feature matrix based on the nonlinear activation function to enhance the model's response to weak fault nonlinear features and suppress background noise interference, generating a high-sparseness deep feature body. S5.5: Perform a final pooling downsampling operation on the high-sparseness deep feature volume to compress spatial redundancy information and preserve the semantic topology of key faults, and finally output the penultimate layer feature map.

[0033] In this embodiment, step S6 involves calculating a gradient heatmap based on the penultimate layer feature map and the final classification layer output, using a gradient-weighted class activation mapping algorithm. This specifically includes the following steps: S6.1: Perform forward propagation processing on the second-to-last layer feature map to obtain the category probability distribution vector of the final classification layer, so as to clarify the confidence basis for the current input vibration signal to be judged as a specific fault type; S6.2: Based on the scalar output value of the target fault category in the category probability distribution vector, perform backpropagation operation to calculate the gradient tensor of the target fault category relative to each spatial location pixel in the penultimate feature map, so as to quantify the rate of change of sensitivity of each local feature region to the fault decision result. S6.3: Perform global average pooling on the gradient tensor along the channel dimension to generate a channel importance weight vector corresponding to the target fault category, so as to extract the global contribution strength index of each feature channel to the identification of the specific fault mode.

[0034] S6.4: Element-wise weighted fusion calculation is performed using the channel importance weight vector and the penultimate layer feature map, and nonlinear activation processing is performed through a linear rectifier function to generate a preliminary fault attention response heatmap, so as to highlight the key areas in the vibration signal time-frequency domain that have a significant positive correlation with fault diagnosis. S6.5: Perform bilinear interpolation upsampling on the preliminary fault attention response heatmap to map its spatial resolution to the same size as the time-frequency plane of the original preprocessed vibration signal, generating the final spatially aligned gradient heatmap to achieve precise geometric matching between the abstract feature attention inside the neural network and the time-frequency coordinates of the original vibration signal.

[0035] In this embodiment, step S7 involves inversely mapping the gradient heatmap to the time-frequency plane via short-time Fourier transform, and then overlaying the physical frequency location labels corresponding to the fault type from the fault semantic anchor point library to generate a joint attribution map. This process specifically includes the following steps: S7.1: Perform short-time Fourier transform inverse transformation on the gradient weighted class activation map heatmap corresponding to the penultimate layer feature map to restore the abstract spatial attention weights to a time-frequency sensitive region mask with time resolution and frequency resolution, and obtain the initial time-frequency sensitivity distribution matrix. S7.2: Based on the discretized sparse vectors of the corresponding fault types in the fault semantic anchor point library, extract the theoretical fault feature frequency values ​​represented by non-zero elements, and perform a frequency band expansion operation on the theoretical fault feature frequency values ​​to generate physical frequency position rectangular labels covering the allowable error range, and obtain the physical prior frequency mask. S7.3: Dynamically adjust the fusion weight coefficients using the output confidence values ​​of the model's final classification layer to balance the contribution ratios of data-driven attention and physical prior knowledge in the final result, thereby obtaining an adaptive dynamic fusion factor. S7.4: Based on the adaptive dynamic fusion factor, a weighted superposition fusion operation is performed on the initial time-frequency sensitivity distribution matrix and the physical prior frequency mask to eliminate the bias of a single perspective and highlight the overlapping high-confidence region, so as to obtain a preliminary fused time-frequency joint heat map. S7.5: Visualize and render the initially fused time-frequency joint heatmap, overlay the highlighted areas onto the envelope spectrum of the original vibration signal and annotate the fault semantic description information to generate the final joint attribution map for engineers to verify the consistency of the diagnostic logic.

[0036] In this embodiment, step S8 determines whether the overlap between the model's region of interest and the physical region in the joint attribution graph meets a preset threshold. If it does not meet the threshold, a physical inconsistency alarm is output. Specifically, this includes the following steps: S8.1: Perform binarization thresholding on the final joint attribution map to extract the highlighted data-driven model attention region mask and the physical prior expected region mask, generating a region set object containing two independent binary matrices; S8.2: Based on the two independent binary matrices in the generated region set object, perform pixel-level logical AND operation and union statistical calculation to quantify the number of overlapping pixels and the total number of covered pixels between the model's focus area and the physical expected area on the time-frequency plane, and generate a spatial overlap ratio parameter. S8.3: Use the preset physical consistency confidence threshold to compare and judge the magnitude of the generated spatial overlap ratio parameter to evaluate the physical mechanism conformity of the current fault diagnosis result and generate a Boolean type logical consistency verification result flag bit; S8.4: For abnormal states where the logical consistency check result flag indicates that the preset threshold is not met, call the alarm interface of the industrial control system and encapsulate error description information containing deviation type and time-frequency coordinates to generate a physical inconsistency alarm command data packet. S8.5: In response to the generated physical inconsistency alarm command data packet, perform the diagnostic result freeze operation and block the output channel of the normal fault classification label. At the same time, push the audit trail record containing the visual difference comparison to the human-machine interface and generate the closed-loop log of the intercepted untrusted diagnosis and handling.

[0037] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0038] Unless otherwise defined, the technical or scientific terms used herein should be understood in their ordinary sense by one of ordinary skill in the art to which this invention pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” involved in the embodiments of this invention refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0039] The above description is merely an exemplary embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for fault prediction of transmission components in a powder packaging machine based on vibration signal analysis, characterized in that, Includes the following steps: S1: Collect the vibration time-domain signal of the transmission components of the powder packaging machine under rated operating conditions and obtain the raw vibration signal data; S2: Perform bandpass filtering and normalization preprocessing on the original vibration signal data to eliminate background noise interference and generate a preprocessed vibration signal; S3: Based on the multi-level dynamic constraints of the transmission components of the powder packaging machine, the theoretical fault feature frequency families of the geared motor, synchronous belt, gearbox, bearing and cam mechanism are extracted, and each frequency family is discretized into a sparse vector to construct a fault semantic anchor point library. S4: Input the preprocessed vibration signal into the convolutional neural network model. During the training phase, calculate the KL divergence loss between the feature vector and the fault semantic anchor library through the frequency domain projection head, and optimize the parameters of the convolutional neural network model in conjunction with the cross-entropy loss to generate a semantic anchor convolutional neural network model. S5: In the fault diagnosis reasoning stage, the preprocessed vibration signal is input into the semantic anchored convolutional neural network model to obtain the penultimate layer feature map; S6: Based on the penultimate layer feature map and the final classification layer output, the gradient weighted class activation mapping algorithm is used to calculate the gradient heatmap; S7: The gradient heatmap is inversely mapped to the time-frequency plane via short-time Fourier transform, and the physical frequency location labels corresponding to the fault type in the fault semantic anchor point library are superimposed to generate a joint attribution map. S8: Determine whether the overlap between the model's region of interest and the physical region in the joint attribution graph meets the preset threshold. If not, output a physical inconsistency alarm.

2. The method for fault prediction of transmission components of a powder packaging machine based on vibration signal analysis according to claim 1, characterized in that, In step S1, the triaxial accelerometer is driven by the acquisition control command to capture the analog mechanical vibration of the transmission component at the rated speed. The analog quantity is quantized into a discrete time series by the analog-to-digital conversion module to generate a single-channel raw vibration digital signal. Timestamp verification and packet loss detection are performed on each single-channel raw vibration digital signal to remove abnormal data segments caused by transmission jitter. The time reference of the multi-channel data is aligned to generate a spatiotemporally aligned multi-source vibration signal set.

3. The method for fault prediction of transmission components of a powder packaging machine based on vibration signal analysis according to claim 1, characterized in that, In step S2, specifically: the time-domain discrete sequence of the multi-source vibration signal set is converted into a frequency-domain power spectral density distribution using the fast Fourier transform algorithm to generate an initial frequency-domain feature vector; Based on the non-fault background noise frequency band boundary of the initial frequency domain feature vector, a Butterworth digital bandpass filter transfer function with steep cutoff characteristics is constructed, and a convolution filtering operation is performed on the original vibration signal data to generate a bandpass filtered vibration signal. For the time-domain waveform envelope of the bandpass filtered vibration signal, the Hilbert transform algorithm is used to extract the instantaneous amplitude sequence, and the global root mean square value is calculated as a dynamic scaling reference to generate a normalized reference scalar. The normalized reference scalar is used to perform element-wise division on each sampling point of the bandpass filtered vibration signal, and the result is mapped to the standard normal distribution interval.

4. The method for fault prediction of transmission components of a powder packaging machine based on vibration signal analysis according to claim 1, characterized in that, In step S4, based on the fault semantic anchor point library and preprocessed vibration signals, a convolutional neural network architecture containing a learnable frequency domain projection head is constructed. The frequency domain projection head is configured to receive the feature vector after global average pooling and output a sparse probability distribution vector to establish a mapping channel between latent space features and physical fault frequency families.

5. The method for fault prediction of transmission components of a powder packaging machine based on vibration signal analysis according to claim 4, characterized in that, In step S4, specifically: Multi-layer convolution and pooling operations are performed on the preprocessed vibration signal to extract deep time-frequency feature maps, and global average pooling is performed on the deep time-frequency feature maps to generate feature vectors to be projected. The frequency domain projection head is used to perform nonlinear activation and element-wise weighting on the feature vector to be projected, generating a model-predicted sparse probability distribution vector, and the physical constraint divergence loss value is calculated based on the model-predicted sparse probability distribution vector and the fault semantic anchor vector corresponding to the current sample. The cross-entropy classification loss value is calculated based on the output of the final classification layer and the real fault label. The physical constraint divergence loss value and the cross-entropy classification loss value are then weighted and fused to generate the joint optimization total loss value. Gradient backpropagation and parameter update are then performed based on the joint optimization total loss value.

6. The method for fault prediction of transmission components of a powder packaging machine based on vibration signal analysis according to claim 1, characterized in that, In step S5, the preprocessed vibration signal is reconstructed using tensor dimensions to generate a standardized input tensor; and a multi-layer convolution kernel sliding weighted summation operation is performed on it to extract local high-frequency impact features and low-frequency modulation modes in the vibration signal, generating a multi-channel primary feature map.

7. The method for fault prediction of transmission components of a powder packaging machine based on vibration signal analysis according to claim 6, characterized in that, Step S5 also includes performing distribution translation and scaling correction on the multi-channel primary feature map using a batch normalization algorithm to obtain an intermediate feature matrix; performing threshold truncation and sparsification on the intermediate feature matrix using a nonlinear activation function to generate a high-sparseness deep feature body; performing a final pooling downsampling operation on the high-sparseness deep feature body to compress spatial redundancy information and retain the key fault semantic topology, and finally outputting the penultimate layer feature map.

8. The method for fault prediction of transmission components of a powder packaging machine based on vibration signal analysis according to claim 1, characterized in that, In step S6, the penultimate layer feature map is forward propagated to obtain the category probability distribution vector of the final classification layer; Based on the scalar output value of the target fault category in the category probability distribution vector, perform backpropagation operation to calculate the gradient tensor of the target fault category relative to each spatial location pixel in the penultimate feature map; Perform a global average pooling operation on the gradient tensor along the channel dimension to generate a channel importance weight vector corresponding to the target fault category; The channel importance weight vector and the penultimate layer feature map are fused element-wise, and nonlinear activation is performed through a linear rectified function to generate a preliminary fault attention response heatmap. Bilinear interpolation upsampling is performed on the preliminary fault attention response heatmap to generate a spatially aligned gradient heatmap.

9. The method for fault prediction of transmission components of a powder packaging machine based on vibration signal analysis according to claim 1, characterized in that, In step S7, the gradient weighted class activation map heatmap corresponding to the penultimate feature map is subjected to short-time Fourier transform inverse transformation to restore the abstract spatial attention weights to a time-frequency sensitive region mask with time resolution and frequency resolution, thereby obtaining the initial time-frequency sensitivity distribution matrix. Based on the discretized sparse vectors corresponding to the fault types in the fault semantic anchor point library, the theoretical fault feature frequency values ​​represented by non-zero elements are extracted, and a frequency band expansion operation is performed on the theoretical fault feature frequency values ​​to generate physical frequency position rectangular labels covering the allowable error range, thereby obtaining the physical prior frequency mask.

10. The method for fault prediction of transmission components of a powder packaging machine based on vibration signal analysis according to claim 9, characterized in that, In step S7, a weighted superposition fusion operation is performed on the initial time-frequency sensitivity distribution matrix and the physical prior frequency mask based on the adaptive dynamic fusion factor to obtain a preliminary fused time-frequency joint heat map; the preliminary fused time-frequency joint heat map is visualized and rendered, and the highlighted areas are superimposed on the envelope spectrum of the original vibration signal and labeled with fault semantic description information to generate the final joint attribution map.

Images