Welding equipment condition monitoring method and device based on artificial intelligence

Abstract

The present invention provides an artificial intelligence-based welding equipment status monitoring method and device, relating to the fields of artificial intelligence and data processing technology. Vibration signals generated by target welding equipment during operation are collected in real time to generate a vibration data stream corresponding to the target welding equipment. The method then determines the corresponding time-frequency energy distribution of the vibration data stream, revealing energy variations in both time and frequency. This method enhances the signal-to-noise ratio, reduces interference from non-correlated frequency bands, and strengthens early fault signals, increasing the model's sensitivity to abnormal events. Furthermore, a deep residual neural network employs a preset dynamic gated attention mechanism to extract rapidly changing features and stable structural features from the data to be identified. This method can distinguish and weight these "rapidly changing features" from "stable structural features," improving the model's sensitivity to abnormal events (such as early fault shocks), thereby enabling accurate equipment status analysis of the welding equipment.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and data processing technology, and in particular to a welding equipment status monitoring method and device based on artificial intelligence. Background Art

[0002] With the continuous improvement of intelligent manufacturing and high-end equipment automation, welding equipment, as a key node in the production line, has a high operational stability that is directly related to weld quality, equipment lifespan, and production safety. However, traditional welding equipment condition monitoring methods mainly rely on manual inspections, intermittent recording, or simple threshold alarm systems.

[0003] The vibration signals generated by welding equipment during operation have complex characteristics and variations. Existing technologies typically use fixed low-pass or band-pass filters to filter these vibration signals. However, this approach cannot automatically adjust its cutoff frequency based on these variations, making it difficult to effectively remove noise or retain useful fault signatures. Furthermore, state changes during welding (for example, from normal operation to failure) are often sudden and unpredictable. Existing technologies suffer from slow recognition, making it difficult to accurately capture the moment these transient changes occur. This leads to a high rate of false positives, which impacts the effectiveness and accuracy of subsequent feature extraction. This is particularly prone to missed and false positives in the early stages of subtle faults and sudden anomalies. Summary of the Invention

[0004] In view of this, the object of the present invention is to provide a welding equipment status monitoring method and device based on artificial intelligence, which can enhance the sensitivity of the model to abnormal events.

[0005] In a first aspect, an embodiment of the present invention provides an artificial intelligence-based welding equipment status monitoring method, wherein the method includes: real-time acquisition of vibration signals generated by a target welding equipment during operation to obtain a vibration data stream corresponding to the target welding equipment; wherein the vibration signals are acquired by a vibration sensor installed at a preset position of the target welding equipment; determining the time-frequency energy distribution corresponding to the vibration data stream based on the acquisition time of the vibration data stream; performing signal filtering processing on the vibration data stream based on the frequency band energy ratio corresponding to the time-frequency energy distribution; and performing feature enhancement processing on the vibration data stream after filtering based on the sudden change signal corresponding to the time-frequency energy distribution; generating data to be identified based on the enhanced vibration data stream, inputting the data to be identified into a preset deep residual neural network, performing fault category analysis on the vibration data stream, and determining a fault category analysis result; wherein the deep residual neural network is used to extract rapidly changing features and stable structural features in the data to be identified using a preset dynamic gated attention mechanism, so as to analyze the fault category of the vibration data stream based on the rapidly changing features and the stable structural features; and determining the fault category corresponding to the target welding equipment based on the fault category analysis result.

[0006] In combination with the first aspect, an embodiment of the present invention provides a first implementation method of the first aspect, wherein the step of determining the time-frequency energy distribution corresponding to the vibration data stream based on the acquisition moment of the vibration data stream includes: performing multi-band decomposition on the time domain signal of the vibration data stream, determining the frequency band components corresponding to the vibration data stream in multiple frequency bands, and obtaining the time-frequency energy distribution of the vibration data stream based on each acquisition moment.

[0007] In combination with the first aspect, an embodiment of the present invention provides a second implementation method of the first aspect, wherein the step of performing signal filtering processing on the vibration data stream based on the frequency band energy ratio corresponding to the time-frequency energy distribution includes: calculating the frequency band energy corresponding to a preset time window based on the time-frequency energy distribution; determining the frequency band energy ratio corresponding to the time-frequency energy distribution according to the frequency band energy; determining the dominant frequency corresponding to the vibration data stream based on the frequency band energy ratio; and filtering the preset frequency band of the vibration data stream according to the dominant frequency.

[0008] In combination with the first aspect, an embodiment of the present invention provides a third implementation of the first aspect, wherein the step of performing feature enhancement processing on the vibration data stream after filtering based on the mutation signal corresponding to the time-frequency energy distribution includes: calculating the local energy entropy of the vibration data stream in the corresponding frequency band according to the time-frequency energy distribution; determining the energy entropy deviation amplitude of the local energy entropy corresponding to a preset threshold, and determining the mutation signal point corresponding to the vibration data stream based on the energy entropy deviation amplitude; segmenting the vibration data stream according to the mutation signal point to obtain multiple signal segmentation intervals; determining the time change rate and frequency change rate corresponding to the frequency band component of each segmentation interval; generating a time-frequency enhancement matrix corresponding to the vibration data stream according to the time change rate and the frequency change rate, so as to perform feature enhancement processing on the vibration data stream.

[0009] In combination with the first aspect, an embodiment of the present invention provides a fourth implementation of the first aspect, wherein the deep residual neural network includes multiple layers of time-frequency convolution modules, and the frequency bandwidth of the time-frequency convolution modules at each layer is different; the data to be identified is input into a preset deep residual neural network, and the vibration data stream is subjected to fault category analysis, and the steps of determining the fault category analysis results include: inputting the data to be identified into a preset deep residual neural network, and extracting target local features of the data to be identified at different time scales and frequency ranges through the multiple layers of time-frequency convolution modules of the deep residual neural network; performing feature fusion on the target local features through a preset dynamic gated attention mechanism to generate a feature fusion vector; mapping the feature fusion vector to the fault category space, and determining the fault category analysis result corresponding to the data to be identified.

[0010] In combination with the first aspect, an embodiment of the present invention provides a fifth implementation of the first aspect, wherein, through a preset dynamic gated attention mechanism, feature fusion is performed on the target local features to generate a feature fusion vector, including: extracting the transient features of the target local features of each level through a convolution operation, and determining the steady-state features of the target local features of each level through a pooling method; through a preset gating unit, the transient features and steady-state features of multiple levels are weighted and spliced ​​according to the spectral energy of the corresponding levels to generate a feature fusion vector mixed feature corresponding to the target local features.

[0011] In combination with the first aspect, an embodiment of the present invention provides a sixth implementation of the first aspect, wherein the method further includes: obtaining pre-stored typical fault operating condition data; extracting the spectrum template corresponding to each type of fault in the typical fault operating condition data, and constructing a spectrum structure library; based on the spectrum structure library, performing time-frequency domain transformation on the spectrum template corresponding to each type of fault, and generating an initial convolution kernel corresponding to the deep residual neural network.

[0012] In combination with the first aspect, an embodiment of the present invention provides a seventh implementation of the first aspect, wherein the method further includes: determining the cross-entropy loss corresponding to the deep residual neural network; calculating the KL divergence of the time domain feature distribution and the frequency domain feature distribution corresponding to the preset training sample set, and measuring the similarity of the time-frequency domain feature distribution corresponding to the training sample set; calculating the total loss function corresponding to the deep residual neural network based on the cross-entropy loss, the similarity of the time-frequency domain feature distribution and the preset time-frequency spectrum smoothness constraint; based on the total loss function, as well as the dynamic gating adjustment factor corresponding to the dynamic gated attention mechanism and the preset frequency change response sensitivity, updating the model parameters of the deep residual neural network.

[0013] In combination with the first aspect, an embodiment of the present invention provides an eighth implementation of the first aspect, wherein the method further includes: generating a fault simulation signal corresponding to a preset training sample set based on a preset plurality of typical frequency components; generating an adversarial disturbance vector corresponding to the preset training sample set based on a preset disturbance step size; adding the adversarial disturbance vector to the fault simulation signal to generate an adversarial sample corresponding to the preset training sample set; training the deep residual neural network based on the adversarial sample; and at the same time, respectively calculating the total loss function of the preset training sample set and the adversarial sample, and updating the model parameters of the deep residual neural network based on the corresponding total loss function.

[0014] In a second aspect, an embodiment of the present invention further provides a welding equipment status monitoring device based on artificial intelligence, wherein the device comprises: a data acquisition module for real-time acquisition of vibration signals generated by the target welding equipment during operation to obtain a vibration data stream corresponding to the target welding equipment; wherein the vibration signal is acquired by a vibration sensor installed at a preset position of the target welding equipment; a calculation module for determining the time-frequency energy distribution corresponding to the vibration data stream based on the acquisition moment of the vibration data stream; a data processing module for performing signal filtering processing on the vibration data stream based on the energy proportion of the frequency band corresponding to the time-frequency energy distribution; and a signal processing module for performing signal filtering processing on the vibration data stream based on the energy proportion of the frequency band corresponding to the time-frequency energy distribution. The method comprises the following steps: a mutation signal is detected, and feature enhancement processing is performed on the vibration data stream after filtering; an execution module is used to generate data to be identified based on the vibration data stream after enhancement processing, input the data to be identified into a preset deep residual neural network, perform fault category analysis on the vibration data stream, and determine the fault category analysis result; wherein, the deep residual neural network is used to extract the fast-changing features and stable structural features in the data to be identified using a preset dynamic gated attention mechanism, so as to analyze the fault category of the vibration data stream according to the fast-changing features and stable structural features; an output module is used to determine the fault category corresponding to the target welding equipment according to the fault category analysis result.

[0015] Embodiments of the present invention provide an artificial intelligence-based welding equipment status monitoring method and device. This method collects vibration signals generated by target welding equipment during operation in real time to generate a vibration data stream corresponding to the target welding equipment. The method then determines the corresponding time-frequency energy distribution of the vibration data stream, revealing energy variations in both the time and frequency dimensions. This method enhances the signal-to-noise ratio, reduces interference from non-correlated frequency bands, and strengthens early fault signals, increasing the model's sensitivity to abnormal events. Furthermore, a deep residual neural network employs a preset dynamic gated attention mechanism to extract rapidly changing features and stable structural features from the data to be identified. This method distinguishes and weights these features, improving the model's sensitivity to abnormal events (such as early fault shocks), thereby enabling accurate equipment status analysis of the welding equipment.

[0016] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or understood by practicing the present invention. The purposes and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0017] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. 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 paying any creative work.

[0019] Figure 1 A flowchart of a welding equipment status monitoring method based on artificial intelligence provided by an embodiment of the present invention;

[0020] Figure 2 A flowchart of another welding equipment status monitoring method based on artificial intelligence provided by an embodiment of the present invention;

[0021] Figure 3 A schematic diagram comparing the noise suppression capabilities corresponding to a filtering process provided by an embodiment of the present invention;

[0022] Figure 4 A schematic diagram comparing the effects of filtering processing provided by an embodiment of the present invention;

[0023] Figure 5 A schematic diagram comparing detection accuracy of different window division strategies provided by an embodiment of the present invention;

[0024] Figure 6 A schematic diagram comparing the training convergence speeds of different network structures provided by an embodiment of the present invention;

[0025] Figure 7 A schematic diagram comparing the effects of an initialization method provided in an embodiment of the present invention;

[0026] Figure 8 A schematic diagram comparing the effects of different training strategies provided by an embodiment of the present invention;

[0027] Figure 9 A schematic structural diagram of a welding equipment status monitoring device based on artificial intelligence provided by an embodiment of the present invention;

[0028] Figure 10 A schematic structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work shall fall within the scope of protection of the present invention.

[0030] An embodiment of the present invention provides a welding equipment status monitoring method and device based on artificial intelligence, which can improve the model's sensitivity to abnormal events.

[0031] To facilitate understanding of this embodiment, a welding equipment status monitoring method based on artificial intelligence disclosed in an embodiment of the present invention is first introduced in detail. Figure 1 The flowchart corresponding to the embodiment of the present invention is shown. Figure 1 , including the following steps:

[0032] Step S102 : collecting vibration signals generated by the target welding equipment during operation in real time to obtain a vibration data stream corresponding to the target welding equipment.

[0033] The embodiment of the present invention monitors the corresponding operating status of the welding equipment based on the vibration signal of the welding equipment, such as normal operation, welding abnormality, equipment looseness, motor failure, trajectory deviation, and other abnormalities. The vibration signal is collected by a vibration sensor installed at a preset position of the target welding equipment, mainly a high-precision vibration sensor installed at a key position of the equipment, such as the end of the welding gun, the drive motor, and the base structure. The working conditions of the welding equipment are diverse, such as the current size, welding gun speed, material type, ambient temperature, etc., which may affect the operating status of the equipment. The embodiment of the present invention captures the vibration signal generated by the welding equipment during operation in real time, realizes continuous monitoring of non-stationary vibration signals, and avoids missing transient abnormalities. Vibration data can be collected in real time at a high sampling rate through a high-speed data acquisition card or an industrial acquisition gateway.

[0034] Step S104 : determining the time-frequency energy distribution corresponding to the vibration data stream based on the acquisition time of the vibration data stream.

[0035] Step S106 , performing signal filtering processing on the vibration data stream based on the frequency band energy proportion corresponding to the time-frequency energy distribution; and performing feature enhancement processing on the vibration data stream after filtering based on the sudden change signal corresponding to the time-frequency energy distribution.

[0036] The embodiments of the present invention analyze the time-frequency characteristics of vibration signals, converting them into a time-frequency domain representation and determining the corresponding time-frequency energy distribution. The vibration signals are further filtered and feature-enhanced to highlight potential fault characteristics and optimize signal quality. A complex time-domain signal can be decomposed into multiple sub-signals in different frequency bands. Each sub-signal in a frequency band contains corresponding frequency components, representing the components of the vibration data stream within a specific frequency range.

[0037] Furthermore, by determining the energy percentage of a specific frequency band relative to the total energy of the entire signal within a specific time period, we can understand the importance of each frequency component in the signal and its changes over time, thereby determining the signal's frequency distribution. Filtering the signal based on frequency band energy percentages can address the dynamic changes in the signal's dominant frequency under different operating conditions, ensuring that effective features are not weakened or noise remains.

[0038] During the operation of industrial equipment, state changes (such as faults and load switching) are often transient or sudden. Embodiments of the present invention enhance vibration signals based on the corresponding sudden change signals. This amplification process can be performed on a small time window before and after the sudden change, enhancing transient responses and early fault signals to effectively identify specific patterns or anomalies in the equipment.

[0039] Step S108 : generating data to be identified based on the enhanced vibration data stream, inputting the data to be identified into a preset deep residual neural network, performing fault category analysis on the vibration data stream, and determining a fault category analysis result.

[0040] Step S110: determining the fault category corresponding to the target welding equipment according to the fault category analysis result.

[0041] The present invention utilizes a deep residual neural network model for state inference. In this embodiment, the deep residual neural network employs a preset dynamic gated attention mechanism to extract rapidly changing features (such as local mutations, high-frequency components, and energy transition regions) and stable structural features (such as periodicity, low-frequency dominant frequencies, and stable energy distribution regions) from the data to be identified. The neural network can also adjust the fusion weights of different modal features in a time-sensitive manner to avoid missing high-frequency, weak faults. Furthermore, by outputting the probability distribution of each state category through the model, the current device state can be determined using a maximum probability decision mechanism. Output categories include, but are not limited to, normal operation, welding anomalies, loose equipment, motor failure, trajectory deviation, and other anomalies.

[0042] On the basis of the above embodiment, the embodiment of the present invention also provides another welding equipment status monitoring method based on artificial intelligence, Figure 2 Shows a flowchart corresponding to an embodiment of the invention, referring to Figure 2 , the method comprises the following steps:

[0043] Step S202 : collecting vibration signals generated by the target welding equipment during operation in real time to obtain a vibration data stream corresponding to the target welding equipment.

[0044] In combination with the above embodiments, the vibration data stream collected by the sensor can be received in real time through the edge computing terminal deployed on the welding equipment.

[0045] Step S204 , performing multi-band decomposition on the time domain signal of the vibration data stream, determining frequency band components corresponding to the vibration data stream in multiple frequency bands, and obtaining the time-frequency energy distribution of the vibration data stream based on each acquisition moment.

[0046] In conjunction with the above embodiments, a Fourier transform (such as a short-time Fourier transform (STFT) or a wavelet transform (WPT) can be used to perform multi-band decomposition, revealing the frequency distribution of a signal at a specific moment or time period. Each frequency component represents the characteristics of the signal within that frequency band, enabling further analysis of the signal's frequency composition. In the vibration signal of a mechanical system, low-frequency components may correspond to the system's primary motion mode, while high-frequency components may reflect vibrations caused by surface roughness or minor defects.

[0047] Step S206 : performing signal filtering processing on the vibration data stream based on the frequency band energy proportion corresponding to the time-frequency energy distribution.

[0048] For the frequency band components of each of the above frequency bands, the embodiment of the present invention also calculates the frequency band energy of the time window to better understand and analyze complex signals. The frequency band components provide information in the frequency dimension, focusing on revealing the frequency structure within the signal; the frequency band energy within the time window adds consideration of the time dimension on this basis, focusing on how these frequency components change over time. Among them, if the energy of the high-frequency components in the sound signal emitted by the machine suddenly increases within a certain period of time, it may indicate the emergence of a new noise source or wear of mechanical parts.

[0049] In specific implementation, step S206 is implemented through the following process:

[0050] 1) Based on the time-frequency energy distribution, calculate the frequency band energy corresponding to the preset time window.

[0051] In specific implementation, the calculation method of frequency band energy is expressed as:

[0052]

[0053] The above formula shows that the energy sum of a certain frequency band component in a specific time period is calculated based on the frequency band component. is the frequency band energy within the sliding time window, is the length of the sliding time window; is a local time point within the time window; t is the current moment; is the time of the original signal in frequency band f The weight.

[0054] 2) According to the frequency band energy, determine the frequency band energy ratio corresponding to the time-frequency energy distribution.

[0055] Specifically, it is calculated using the following formula:

[0056]

[0057] in, is the energy proportion of frequency band f at time t; Frequency band The energy at time t.

[0058] 3) Based on the energy proportion of the frequency band, determine the dominant frequency corresponding to the vibration data stream.

[0059] The dominant frequency refers to the frequency component with the strongest energy or the greatest contribution within a certain period of time, representing the main oscillation mode or characteristic frequency of the signal in the current state. The calculation method is expressed as:

[0060]

[0061] in, is the current dominant frequency.

[0062] 4) Filter the preset frequency band of the vibration data stream according to the dominant frequency.

[0063] Vibration sensor data is subject to high-frequency noise interference and a wide range of amplitude fluctuations under different operating conditions. Conventional technologies typically use low-pass filtering or normalization with a fixed cutoff frequency, which cannot cope with the dynamic changes in the dominant frequency of the signal under different operating conditions, resulting in weakening of effective features or residual noise. When fault information exists in overlapping frequency bands, filtering performance is significantly degraded. Furthermore, existing technologies, through global normalization, weaken local abnormal vibration characteristics.

[0064] In combination with the above steps, the present invention adopts a dynamic adaptive filtering function to analyze the local time-frequency energy distribution at each moment, perform multi-band decomposition on the original monitoring signal of the vibration sensor, extract the components of each frequency band, and reconstruct it according to its local energy weight. Specifically, at each moment, the energy proportion of each frequency band is calculated, and the signal components of the corresponding frequency band are weighted and fused based on this. At the same time, the adaptive frequency band weight function is calculated in combination with the current dominant frequency, so that the signal near the dominant frequency has a higher weight, thereby forming the final filtered output signal, effectively removing the low-energy noise band and retaining the main vibration information, and then obtaining the filtered data, which is expressed as:

[0065]

[0066] Where, is the filtered data of the i-th sample at time t; is the first frequency band index, is the second frequency band index, and ; is the total number of frequency bands. is the fth frequency band component obtained by STFT of the original signal; is the adaptive band weight, and its calculation method is expressed as: ; It is a bandwidth adjustment factor, which is used to dynamically adjust the weights of signals in different frequency bands during the processing process, thereby achieving accurate signal filtering and feature enhancement.

[0067] Step S208 : performing feature enhancement processing on the vibration data stream after filtering based on the sudden change signal corresponding to the time-frequency energy distribution.

[0068] 1) Based on the time-frequency energy distribution, calculate the local energy entropy of the vibration data stream in the corresponding frequency band.

[0069] At each time point, the proportion of each frequency band's energy in the total energy is first calculated, and then the corresponding energy entropy value is calculated based on the probability structure of the frequency distribution. The magnitude of the energy entropy reflects the degree of concentration or dispersion of the spectrum distribution at the current moment, and is used to characterize the complexity of the change of the current spectrum state, which is expressed as:

[0070]

[0071] Where, is the local energy entropy at time t; is the energy proportion of frequency band f at time t; It is a logarithmic function, with base 10 by default.

[0072] 2) Determine the energy entropy deviation amplitude of the local energy entropy corresponding to a preset threshold, and determine the mutation signal point corresponding to the vibration data stream based on the energy entropy deviation amplitude.

[0073] In a continuous time window, the deviation between the current energy entropy and the mean of the historical window is calculated and compared with the standard deviation threshold. If the deviation exceeds the set range, the moment is determined to be a mutation point, effectively capturing potential faults or state variation behaviors in the signal, thereby automatically delineating signal segments.

[0074] In the embodiment of the present invention, when When , it is marked as a segmentation point, L is the length of the historical window, For time The local energy entropy, is the standard deviation of the entropy value of the historical window.

[0075] 3) The vibration data stream is segmented according to the mutation signal points to obtain multiple signal segmentation intervals; and the time change rate and frequency change rate corresponding to the frequency band component of each segmentation interval are determined.

[0076] 4) Generate a time-frequency enhancement matrix corresponding to the vibration data stream according to the time change rate and the frequency change rate to perform feature enhancement processing on the vibration data stream.

[0077] In its specific implementation, the embodiment of the present invention adopts a dynamic segmentation strategy based on energy mutation detection. Based on the energy distribution characteristics of the signal in the time-frequency domain, it dynamically detects local structural mutations and determines segmentation points that are more consistent with the characteristics of the physical process. The signal is then windowed within each segment interval to achieve stable extraction of local features. Based on the above steps, the starting and ending points of each segment window are determined by the change in local energy entropy. Furthermore, the embodiment of the present invention establishes a weighted matrix for feature extraction within each segment to enhance the expressive power and recognition accuracy of the subsequent model.

[0078] Conventional sliding window methods, however, ignore the sudden changes in the vibration signal's temporal patterns. This makes it impossible to accurately align the start and end points of device state changes within the signal segment, which can easily cause the feature extraction window to overlap multiple state segments. This mismatch between the feature extraction window and the physical process prevents the model from learning pure state features during training, impacting classification accuracy.

[0079] Specifically, within each determined segmented interval, the embodiment of the present invention calculates the temporal rate of change and the rate of change of the frequency weight of the signal segment. The two are then combined by tensor product to generate a two-dimensional enhancement matrix. The two-dimensional enhancement matrix can couple the temporal trend and the frequency domain feature changes, enhancing the joint representation of local complex patterns and key features in the time-frequency domain. It is expressed as:

[0080]

[0081] Where, is the time-frequency enhancement matrix of the i-th sample in the segmented interval; is the gradient of the filtered signal in the time dimension; is the time differential operation; is the tensor product operation; is the gradient of the adaptive band weight in the frequency domain dimension; is the frequency domain differential operation; is the starting time of the segment; The end time of the segment.

[0082] In step S210, based on the enhanced vibration data stream, data to be identified is generated; the data to be identified is input into a preset deep residual neural network, and the target local features of the data to be identified at different time scales and frequency ranges are extracted through the time-frequency convolution modules of multiple levels of the deep residual neural network.

[0083] In one embodiment, the structure of the deep residual neural network adopted in the embodiment of the present invention is composed of 12 layers of time-frequency convolution modules and 4 levels of skip connections: the input layer receives the time-frequency enhancement matrix, which is processed by 3 residual blocks, each residual block contains a cascaded void convolution group, batch normalization and ELU activation function; the skip connection adopts a dynamic gated attention mechanism to generate gating weights through transient-steady-state dual-path feature fusion; the 4th layer of cascaded multi-scale feature fusion layer splices the feature maps of different receptive fields weighted by spectral energy; the final output layer uses the softmax function to map to the fault category space.

[0084] Traditional convolution or spectrum graph input methods only construct features in the time or frequency dimension, making it difficult to capture the dynamic interaction between the two, and lack the ability to express local time-frequency coupling patterns, resulting in insufficient recognition of sudden faults by the model. In order to solve the coupling problem of multi-scale fault features in vibration signals, the present invention adopts a hierarchical increasing hole convolution structure, that is, a multi-level time-frequency convolution module. Among them, different levels use hole convolution operations with different expansion rates to extract local features corresponding to different time scales and frequency ranges in the vibration signal, respectively, to form a multi-scale fusion framework. Each convolution kernel limits the frequency range of its concern in the frequency domain through bandwidth constraints, so that the features at all levels have non-overlapping frequency, improving the spectral resolution and hierarchical structure perception ability of the overall model, which is expressed as:

[0085]

[0086] Where, is the feature map of layer l; d is the index of the dilated convolution layer; D is the index of the dilated convolution layer; is the convolution kernel of the dth level in the lth layer; The void rate is The dilated convolution, Set according to the exponential growth rule, such as, ; is the feature map of the l-1th layer.

[0087] Furthermore, the bandwidth of each convolutional layer is constrained by bandwidth, forcing each layer to focus on frequency bands with different center frequencies, forming a progressive multi-scale analysis, which can be expressed as:

[0088]

[0089] Where, is the frequency domain response of the d-th level convolution kernel; is the L1 norm; is the d-th level bandwidth constraint coefficient; is a frequency variable; is the center frequency of the dth level; is the standard deviation of the d-th level bandwidth; is an exponential function with natural numbers as base.

[0090] In step S212, the target local features are fused through a preset dynamic gated attention mechanism to generate a feature fusion vector.

[0091] In specific implementation, the embodiment of the present invention extracts the transient features of the target local features at each level through convolution operations, and determines the steady-state features of the target local features at each level through a pooling method; through a preset gating unit, the transient features and steady-state features of multiple levels are weighted and spliced ​​according to the spectral energy of the corresponding levels to generate a feature fusion vector corresponding to the target local features.

[0092] Specifically, the traditional attention mechanism focuses on global static feature extraction, and has weak recognition capabilities for sudden and short-term vibration anomalies. It is difficult to capture the correlation between transient features and steady-state features in vibration signals, and is prone to missing high-frequency weak faults. The present invention adopts a dual-modal gating unit. In the feature fusion stage, a gating unit jointly driven by transient features and steady-state features is used to extract rapidly changing features in the input signal through convolution operations, and at the same time use a pooling method to obtain its stable structural features. The fusion ratio between the two is then dynamically adjusted through the gating function to obtain a more discriminative mixed feature. Among them, the data of each frequency band can be mixed, and then multiple frequency bands can be spliced ​​to determine the corresponding mixed features (that is, feature fusion vectors). The gating coefficient changes with time and can be trained, so that the model has the ability to adaptively select important modes, thereby realizing dynamic gated attention enhancement, which is expressed as:

[0093]

[0094]

[0095] Where, is the fused feature vector; is the Sigmoid activation function; is the gating coefficient, which dynamically adjusts the fusion ratio of the two types of features; is element-wise multiplication; is a transient feature, expressed as ; It is a one-dimensional convolution operation; is the transient convolution kernel size; is the steady-state characteristic, expressed as ; is the average pooling operation; is the pooling window length.

[0096] Step S214 : Map the feature fusion vector to the fault category space to determine the fault category analysis result corresponding to the data to be identified.

[0097] Step S216: Determine the fault category corresponding to the target welding equipment according to the fault category analysis result.

[0098] The output fault category analysis results are achieved by pre-labeling the training sample set. The data labeling can rely on the equipment operation log, expert manual judgment and actual fault cases. By synchronously comparing the vibration signal and the operating status, the collected data is labeled at the event level and time period level. The labeling system has an embedded dynamic visualization interface, which combines spectrum analysis and time domain trend analysis to assist expert decision-making. Corresponding to the above embodiment, the labeling categories include but are not limited to "normal operation", "welding abnormality", "loose equipment", "motor failure", "trajectory deviation", "other abnormalities", etc. The labeling results are bound to the vibration data to form a labeled training set, which can be used for subsequent model supervision training.

[0099] In summary, another method for monitoring the state of welding equipment based on artificial intelligence is provided in an embodiment of the present invention. It adopts a filtering mechanism based on local time-frequency energy weights and dominant frequency guidance to construct a moment-adaptive frequency band weighting model. It can dynamically adjust the signal retention ratio of different frequency bands, significantly enhance the dominant feature retention capability and remove overlapping noise. In order to verify the signal processing capability of the dynamic adaptive filtering method under different noise interference, the embodiment of the present invention also provides a comparative diagram of the noise suppression capability corresponding to the filtering processing (refer to Figure 3 ) and filter effect comparison diagram (refer to Figure 4), the filtering method of the embodiment of the present invention is compared with the traditional fixed cutoff frequency low-pass filter and wavelet threshold denoising method. In the experiment, the horizontal axis represents the noise level coefficient, and the larger the value, the stronger the environmental interference; the vertical axis is the signal-to-noise ratio, which reflects the intensity ratio of the effective signal to the noise. The results show that as the noise level increases, the signal-to-noise ratio of the traditional method shows a significant downward trend, while the method of the present invention maintains stable signal quality in high-noise scenarios through dynamic frequency band energy weight adjustment. The fluctuations in the curve represent the random changes in the vibration amplitude of the equipment under actual working conditions. The slope of the broken line of the method of the present invention is the smallest and is always at the top, indicating that it can adaptively suppress the noise overlapping with the effective frequency band according to the real-time frequency domain characteristics, avoid the weakening of local abnormal characteristics by global normalization, and verify the robustness advantage of dynamic filtering in complex industrial environments.

[0100] Different from the traditional fixed sliding window method, the present invention adopts a segmented detection strategy driven by energy entropy changes, automatically identifies the mutation points of the signal, accurately aligns the physical behavior boundaries, improves the matching degree between the local window and the actual working conditions, and effectively improves the temporal stability of the fault characteristics. In each dynamic segmentation interval, the present invention jointly considers the time gradient and the frequency domain weight change rate to form an enhanced matrix of a two-dimensional tensor structure, strengthens the time-frequency coupling expression capability, and enables the deep model to simultaneously capture transient changes and spectral morphological characteristics. The embodiment of the present invention also provides a schematic diagram for comparing detection accuracy of different window division strategies, refer to Figure 5 , compare the effects of different window division strategies on fault detection accuracy, including fixed-length sliding windows, mutation detection methods based on statistical standard deviations, and the energy entropy dynamic segmentation strategy of the present invention. The horizontal axis of the bar graph represents four typical fault types, and the vertical axis represents the detection accuracy, reflecting the ability of different methods to identify fault characteristics. The column shape of the method of the present invention is significantly higher than that of other methods in all fault categories, especially in the "trajectory deviation" scenario with frequent mutations. The texture difference of the column in the figure is used to distinguish different methods. The dot texture column of the dynamic segmentation is highly stable, indicating that it can accurately capture the structural changes in the time-frequency domain of the signal through energy entropy mutation detection, so that the windowing operation is synchronized with the actual state switching of the physical process of the equipment, thereby improving the model's perception sensitivity to transient abnormal events and reducing the problem of feature omission caused by window mismatch.

[0101] Furthermore, the embodiments of the present invention also construct a dynamic gating mechanism for transient-steady-state fusion, and adjust the fusion weights of different modal features in a time-sensitive manner in the residual block, thereby significantly improving the model's attention selection ability in identifying weak anomalies and non-steady-state fault features.

[0102] In order to analyze the training efficiency of different network architectures, the embodiment of the present invention also provides a schematic diagram for comparing the training convergence speed of different network structures. Figure 6, compared with the traditional convolutional neural network, the standard residual network and the multi-scale time-frequency convolution structure of the present invention. The horizontal axis of the line graph is the number of training rounds, the vertical axis is the loss value, and the rate of decline of the curve reflects the speed of model convergence. Due to the lack of frequency band constraints, the loss of traditional convolutional networks decreases slowly and fluctuates greatly in the later period; the curve of the method of the present invention decreases rapidly in the early stage and remains stable, thanks to the convolution kernel initialization strategy based on the fault spectrum and the hierarchical increasing hole convolution design. The shape of the marker symbols in the figure distinguishes different networks. The star-shaped marker trajectory of the method of the present invention is closest to the horizontal axis, indicating that it can quickly extract the key fault-related modes in the vibration signal through the fusion of multi-scale frequency domain features, reduce redundant calculations, and verify the learning efficiency advantage of the network structure for time-frequency coupling features.

[0103] Furthermore, the embodiment of the present invention also designs the initialization method of the model. Most deep models use general initialization methods such as random or Xavier / He, which do not take into account the spectral structure characteristics of welding vibration data and are difficult to match the time-frequency structure characteristics of the vibration signal, resulting in low efficiency of the early training process, slow convolution kernel convergence process, and the model is insensitive to key frequency band responses.

[0104] The present invention adopts a kernel initialization strategy based on the spectrum of typical fault modes. In its specific implementation, it obtains pre-stored typical fault condition data; extracts the spectrum template corresponding to each type of fault from the typical fault condition data, and constructs a spectrum structure library; based on the spectrum structure library, performs a time-frequency domain transformation on the spectrum template corresponding to each type of fault to generate the initial convolution kernel corresponding to the deep residual neural network. Specifically, known typical fault condition data is first collected, and the spectrum template corresponding to each type of fault is extracted through spectrum analysis methods to form a spectrum structure library. Each template represents the characteristic response of a certain type of fault in the frequency domain and serves as a reference for the design of the initial convolution kernel to enhance the model's prior perception capability, expressed as: , is the reference spectrum template of the kth fault, k is the fault category index, and K is the total number of fault categories.

[0105] Furthermore, a weighted combination of each type of fault spectrum template and a preset frequency band response function is performed, and initial convolution kernels are generated through frequency domain to time domain transformation. These convolution kernels have targeted frequency band response capabilities in their initial state. Different fault modes are coupled between the convolution kernels through adjustable parameters to achieve guided training with physical interpretability, which is expressed as:

[0106]

[0107] Where, is the initialization weight of the mth convolution kernel; is the trainable coupling coefficient; is the inverse Fourier transform; A function is used to select the preset frequency band for the mth core. For example, the frequency band with the largest information entropy is selected based on information entropy.

[0108] This embodiment of the present invention uses a fault spectrum template-based initialization method, leveraging the frequency domain responses of known typical fault modes as prior knowledge to design physically interpretable convolution kernel initialization weights. This can fundamentally improve the model's ability to perceive critical frequency bands and accelerate model convergence.

[0109] The embodiment of the present invention also shows a schematic diagram of the effect comparison of the initialization method. Figure 7 By using the kernel density map of the feature space distribution, the physical guidance effect of the spectrum template initialization strategy on the deep model feature learning process is analyzed. Compared with traditional random initialization, the features generated by the spectrum guidance method show clearer category boundaries and more compact intra-class distribution in the dimensionality reduction space. The randomly initialized feature clouds in the left figure have a large number of overlapping areas, indicating that the model has difficulty distinguishing the essential features; the spectrum-guided feature distribution in the right figure forms spatially isolated clusters, the distance between the centers of each category is significantly expanded, and the distribution shape is highly correlated with the physical mode of the fault. The essential improvement of the inter-class separability in the feature space verifies that the convolution kernel initialization strategy based on prior knowledge of the fault spectrum can guide the network to quickly establish a representation space that conforms to the essential characteristics of the vibration signal.

[0110] Furthermore, to improve the ability to identify occasional faults, an embodiment of the present invention also designs a loss function in the form of a polynomial combination, and updates the model parameters based on this loss function. In specific implementation, the cross-entropy loss corresponding to the deep residual neural network is determined; the KL divergence of the time-domain feature distribution and the frequency-domain feature distribution corresponding to the preset training sample set is calculated to measure the similarity of the time-frequency domain feature distribution corresponding to the training sample set; and the total loss function corresponding to the deep residual neural network is calculated based on the cross-entropy loss, the similarity of the time-frequency domain feature distribution, and the preset time-frequency spectrum smoothness constraint.

[0111] Specifically, the calculation method of the total loss function is expressed as:

[0112]

[0113] Where L is the total loss function; the first term is the conventional cross entropy loss, which is used to optimize the classification accuracy is the cross entropy loss, is the cross entropy loss weight. The second term is used to constrain the continuity and smoothness of the model output in the time dimension to prevent abnormal spike interference. The third term measures the similarity between the time domain feature distribution and the frequency domain feature distribution, forcing the two to maintain consistency in a statistical sense, reducing inter-modal conflicts and improving the model's adaptability to complex working conditions. is the KL divergence loss weight. is the second-order derivative of time; is the second-order time derivative of the time spectrum; is the L2 norm; m is the characteristic mode index; M is the total number of modes; is the KL divergence; is the time domain feature distribution; is the frequency domain feature distribution.

[0114] Furthermore, based on the total loss function, the dynamic gating adjustment factor corresponding to the dynamic gated attention mechanism, and the preset frequency change response sensitivity, the model parameters of the deep residual neural network are updated. Traditional stochastic gradient descent or Adam optimizers are prone to falling into local optimality when learning complex time-frequency structures. The present invention adopts a gradient weighted update strategy based on multimodal collaborative attention, combined with the gradient response of transient-steady-state gating features, dynamically adjusts the parameter update rate of each layer, and combines it with the frequency band sensitivity adjustment factor to achieve controllable parameter update and frequency adaptability, expressed as:

[0115]

[0116] Where, is the model parameter of the lth layer at the tth iteration; is the model parameter of the lth layer at the t+1th iteration; For the basic learning rate, for example, set it to 0.001. is the dynamic gating adjustment factor of the lth layer at the tth iteration, which indicates the importance weight of the current focus mode of the layer. It can adaptively control the learning rate of different modal features, especially give differentiated learning updates to the residual module with frequency change sensitivity, thereby significantly improving the training efficiency and model convergence quality. It is expressed as:

[0117]

[0118] In the above formula, is the gated output of layer l at iteration t; is the transient feature of the layer l at the tth iteration; For the Layer The steady-state characteristics of the layer at the iteration; is the gradient of the total loss function with respect to the parameters of the lth layer. It is the frequency band sensitivity adjustment factor, which indicates the response sensitivity of the l-th layer convolution kernel to frequency changes and is preset manually. is the modal collaborative difference term of the lth layer, defined as: ; is the L2 norm; Indicates the calculation of the mean.

[0119] Furthermore, existing model training mostly relies on supervised training of existing labeled samples, and has weak recognition capabilities when facing unseen fault modes. In addition, it usually lacks regularization methods based on adversarial perturbations or synthetic abnormal patterns, making the model prone to misjudging out-of-distribution data. To enhance the model's generalization ability for unseen fault modes, an embodiment of the present invention also generates a fault simulation signal corresponding to a preset training sample set based on multiple preset typical frequency components; generates an adversarial perturbation vector corresponding to the preset training sample set based on a preset perturbation step size; adds the adversarial perturbation vector to the fault simulation signal to generate an adversarial sample corresponding to the preset training sample set; trains a deep residual neural network based on the adversarial sample; and simultaneously calculates the total loss function of the preset training sample set and the adversarial sample, and updates the model parameters of the deep residual neural network based on the corresponding total loss function.

[0120] Specifically, the embodiment of the present invention enhances the model's ability to learn potential fault features by injecting simulated fault features during the training process. The fault simulation signal is formed by the superposition of multiple typical frequency components, with a representative amplitude attenuation form, simulating the propagation process after the occurrence of a real fault. The superposition can be performed by adding corresponding element positions. At the same time, adversarial sample perturbations are added to the feature space, and the original samples and perturbed samples are jointly optimized through the loss function, forcing the model to learn a more robust and generalized feature boundary structure, thereby effectively identifying unseen abnormal patterns. First, the vibration signal is enhanced, which is expressed as:

[0121]

[0122] Where, is the enhanced vibration signal (also known as the fault simulation signal); is the original signal; is the disturbance intensity coefficient; is the failure mode index; is the total number of failure modes; For the kth x Amplitude coefficient of class fault; is the attenuation factor; t is the time variable; For the kth x The characteristic frequency of the fault type.

[0123] Then, adversarial perturbation is implemented in the feature space, and the decision boundary robustness is improved by alternately optimizing the loss of the original sample and the adversarial sample, which can be expressed as:

[0124]

[0125] Where, To counter the disturbance vector; is the perturbation step length; is the input gradient; is the loss function; Feature representation for adversarial samples; is the target label.

[0126] Furthermore, the generated perturbation vector is superimposed on the enhanced vibration signal (i.e., the fault simulation signal) to obtain an adversarial sample. During training, the total loss function of the original sample and the adversarial sample is calculated simultaneously. The model parameters are updated through backpropagation. The model is required to minimize the loss of both, forcing the feature extraction module to ignore the perturbation noise and focus on the essential fault characteristics. In the network training process of this embodiment of the present invention, if the validation set loss does not decrease significantly over several consecutive rounds, such as a decrease of less than 0.001 over five consecutive rounds, or the number of training rounds exceeds a preset maximum number of rounds, such as 200 rounds, the training is determined to have converged and the training process is terminated early to avoid overfitting.

[0127] The embodiment of the present invention simulates the real fault propagation process through frequency construction and amplitude attenuation characteristics, and jointly adds anti-disturbance samples into the training process to enhance the model's generalization ability and robustness to new types of faults that have never been seen, breaking through the problem of traditional models' dependence on a single sample.

[0128] In order to evaluate the impact of different training strategies on model performance, the embodiment of the present invention also provides a schematic diagram for comparing the effects of different training strategies. Figure 8 , comparing the traditional adaptive moment estimation optimizer with the multimodal gradient weighted update method of the present invention. In the dual-axis line graph, the left vertical axis is the loss value, the right side is the classification accuracy, and the horizontal axis is the number of training rounds. The loss curve of the traditional method decreases gently and the accuracy rises laggingly, while the loss and accuracy curves of the method of the present invention show a trend of coordinated rapid optimization. The solid line in the figure represents the method of the present invention, the dotted line is the comparison method, and the blue and red colors correspond to the loss and accuracy axes, respectively. The method of the present invention distinguishes the gradient contributions of transient and steady-state features through dynamic gating adjustment factors, so that the parameter update is more focused on the fault-sensitive frequency band, thereby achieving lower loss values ​​and higher accuracy under the same number of training rounds, proving that the gradient weighted strategy can effectively balance modal differences and enhance the model's generalization ability for complex vibration modes.

[0129] Furthermore, based on the above embodiment, the embodiment of the present invention also provides a welding equipment status monitoring device based on artificial intelligence, Figure 9 The schematic diagram of the structure corresponding to the embodiment of the present invention is shown. Figure 9The device includes: a data acquisition module 10 for real-time acquisition of vibration signals generated by a target welding device during operation to obtain a vibration data stream corresponding to the target welding device; wherein the vibration signals are acquired by a vibration sensor installed at a preset position of the target welding device; a calculation module 20 for determining a time-frequency energy distribution corresponding to the vibration data stream based on the acquisition time of the vibration data stream; a data processing module 30 for performing signal filtering on the vibration data stream based on the energy proportion of a frequency band corresponding to the time-frequency energy distribution; and performing feature enhancement on the filtered vibration data stream based on a sudden change signal corresponding to the time-frequency energy distribution; an execution module 40 for generating data to be identified based on the enhanced vibration data stream, inputting the data to be identified into a preset deep residual neural network, performing fault category analysis on the vibration data stream, and determining a fault category analysis result; wherein the deep residual neural network is used to extract fast-changing features and stable structural features from the data to be identified using a preset dynamic gated attention mechanism, so as to analyze the fault category of the vibration data stream based on the fast-changing features and stable structural features; and an output module 50 for determining the fault category corresponding to the target welding device based on the fault category analysis result.

[0130] An embodiment of the present invention provides an artificial intelligence-based welding equipment status monitoring device, the implementation principle and technical effects of which are the same as those of the aforementioned method embodiment. For the sake of brief description, for matters not mentioned in the device embodiment, reference can be made to the corresponding content in the aforementioned method embodiment.

[0131] An embodiment of the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned Figures 1 to 2 The embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to execute the above Figures 1 to 2 The embodiment of the present invention also provides a structural diagram of an electronic device, such as Figure 10 FIG. 1 is a schematic diagram of the structure of the electronic device, wherein the electronic device includes a processor 101 and a memory 100, the memory 100 stores computer executable instructions that can be executed by the processor 101, and the processor 101 executes the computer executable instructions to implement the above Figures 1 to 2 Either of the methods shown. Figure 10 In the illustrated embodiment, the electronic device further includes a bus 102 and a communication interface 103 , wherein the processor 101 , the communication interface 103 and the memory 100 are connected via the bus 102 .

[0132] Among them, the memory 100 may include high-speed random access memory (RAM), and may also include non-volatile memory (non-volatile memory), such as at least one disk storage. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 103 (which can be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. can be used. The bus 102 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc., and can also be an AMBA (Advanced Microcontroller Bus Architecture, on-chip bus standard) bus, wherein AMBA defines three types of buses, including APB (Advanced Peripheral Bus) bus, AHB (Advanced High-performance Bus) bus and AXI (Advanced eXtensible Interface) bus. The bus 102 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 10 Only one bidirectional arrow is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0133] The processor 101 may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed by an integrated logic circuit of hardware in the processor 101 or by instructions in the form of software. The above-mentioned processor 101 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in conjunction with the embodiments of the present application can be directly embodied as being executed by a hardware decoding processor, or can be executed by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium mature in the art, such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor 101 reads the information in the memory and combines its hardware to complete the above Figures 1 to 2 Any of the methods shown.

[0134] The embodiments of the present invention provide a computer program product for an artificial intelligence-based welding equipment condition monitoring method and apparatus, comprising a computer-readable storage medium storing program code. The program code includes instructions that can be used to execute the methods described in the aforementioned method embodiments. For specific implementations, please refer to the method embodiments and will not be described in detail here. Those skilled in the art will clearly understand that, for ease and brevity of description, the specific operating processes of the system described above can refer to the corresponding processes in the aforementioned method embodiments and will not be described in detail here. Furthermore, in the description of the embodiments of the present invention, unless otherwise specified or limited, the terms "installed," "connected," and "connected" should be interpreted broadly, for example, to mean fixed, removable, or integral; mechanical or electrical; direct, indirect via an intermediary, or internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances. If the functions described are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or the portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0135] In the description of the present invention, it should be noted that the terms "first", "second" and "third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. Finally, it should be noted that the above embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that any person skilled in the art who is familiar with this technical field can still modify the technical solutions described in the aforementioned embodiments within the technical scope disclosed by the present invention, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A welding equipment status monitoring method based on artificial intelligence, characterized in that: The method comprises: The vibration signal generated by the target welding equipment during operation is collected in real time to obtain a vibration data stream corresponding to the target welding equipment; wherein the vibration signal is collected by a vibration sensor installed at a preset position of the target welding equipment; determining a time-frequency energy distribution corresponding to the vibration data stream based on a collection time of the vibration data stream; performing signal filtering processing on the vibration data stream based on the frequency band energy proportion corresponding to the time-frequency energy distribution; and performing feature enhancement processing on the vibration data stream after filtering based on the mutation signal corresponding to the time-frequency energy distribution; Based on the enhanced vibration data stream, generating data to be identified, inputting the data to be identified into a preset deep residual neural network, performing fault category analysis on the vibration data stream, and determining a fault category analysis result; wherein the deep residual neural network is used to extract rapidly changing features and stable structural features in the data to be identified using a preset dynamic gated attention mechanism, so as to analyze the fault category of the vibration data stream based on the rapidly changing features and the stable structural features; Determining a fault category corresponding to the target welding equipment according to the fault category analysis result; The step of performing signal filtering processing on the vibration data stream based on the frequency band energy proportion corresponding to the time-frequency energy distribution includes: Based on the time-frequency energy distribution, calculating the frequency band energy corresponding to the preset time window; Determining a frequency band energy ratio corresponding to the time-frequency energy distribution according to the frequency band energy; Determining a dominant frequency corresponding to the vibration data stream based on the energy proportion of the frequency band; performing filtering processing on a preset frequency band of the vibration data stream according to the dominant frequency; The deep residual neural network includes multiple levels of time-frequency convolution modules, and the frequency bandwidth of the time-frequency convolution modules at each level is different. The steps of inputting the data to be identified into a preset deep residual neural network, performing fault category analysis on the vibration data stream, and determining the fault category analysis result include: Inputting the data to be identified into a preset deep residual neural network, and extracting target local features of the data to be identified at different time scales and frequency ranges through the multiple layers of time-frequency convolution modules of the deep residual neural network; Perform feature fusion on the target local features through a preset dynamic gated attention mechanism to generate a feature fusion vector; Mapping the feature fusion vector to the fault category space to determine the fault category analysis result corresponding to the data to be identified; The step of performing feature fusion on the target local features through a preset dynamic gated attention mechanism to generate a feature fusion vector includes: Extracting transient features of the target local features at each level through a convolution operation, and determining steady-state features of the target local features at each level through a pooling method; Through a preset gating unit, the transient features and the steady-state features of multiple levels are weightedly spliced ​​according to the spectrum energy of the corresponding levels to generate a feature fusion vector mixed feature corresponding to the target local feature.

2. The method according to claim 1, characterized in that The step of determining the time-frequency energy distribution corresponding to the vibration data stream based on the acquisition time of the vibration data stream includes: The time domain signal of the vibration data stream is subjected to multi-band decomposition to determine frequency band components corresponding to the vibration data stream in multiple frequency bands, thereby obtaining a time-frequency energy distribution of the vibration data stream based on each acquisition moment.

3. The method according to claim 2, characterized in that The step of performing feature enhancement processing on the filtered vibration data stream based on the sudden change signal corresponding to the time-frequency energy distribution includes: Calculating the local energy entropy of the vibration data stream in the corresponding frequency band according to the time-frequency energy distribution; determining an energy entropy deviation amplitude of the local energy entropy corresponding to a preset threshold, and determining a mutation signal point corresponding to the vibration data stream based on the energy entropy deviation amplitude; Segmenting the vibration data stream into intervals according to the sudden signal points to obtain a plurality of signal segment intervals; determining the time change rate and frequency change rate corresponding to the frequency band component of each segment interval; A time-frequency enhancement matrix corresponding to the vibration data stream is generated according to the time change rate and the frequency change rate, so as to perform feature enhancement processing on the vibration data stream.

4. The method according to claim 1, wherein The method further comprises: Obtain pre-stored typical fault condition data; Extracting the spectrum template corresponding to each type of fault in the typical fault condition data and constructing a spectrum structure library; Based on the spectrum structure library, the spectrum template corresponding to each type of fault is transformed in the time-frequency domain to generate an initial convolution kernel corresponding to the deep residual neural network.

5. The method according to claim 1, wherein The method further comprises: Determining a cross entropy loss corresponding to the deep residual neural network; Calculate the KL divergence of the time domain feature distribution and the frequency domain feature distribution corresponding to the preset training sample set to measure the similarity of the time and frequency domain feature distribution corresponding to the training sample set; Calculating a total loss function corresponding to the deep residual neural network based on the cross entropy loss, the time-frequency domain feature distribution similarity, and a preset time-frequency spectrum smoothness constraint; Based on the total loss function, the dynamic gating adjustment factor corresponding to the dynamic gating attention mechanism and the preset frequency change response sensitivity, the model parameters of the deep residual neural network are updated.

6. The method according to claim 5, characterized in that The method further comprises: Generating a fault simulation signal corresponding to the preset training sample set according to a plurality of preset typical frequency components; Based on a preset perturbation step size, generating an adversarial perturbation vector corresponding to the preset training sample set; Adding the adversarial disturbance vector to the fault simulation signal to generate an adversarial sample corresponding to the preset training sample set; Based on the adversarial sample, the deep residual neural network is trained; at the same time, the total loss function of the preset training sample set and the adversarial sample is calculated respectively, and the model parameters of the deep residual neural network are updated based on the corresponding total loss function.

7. A welding equipment status monitoring device based on artificial intelligence, characterized in that: The device comprises: A data acquisition module is configured to acquire vibration signals generated by a target welding device during operation in real time to obtain a vibration data stream corresponding to the target welding device; wherein the vibration signals are acquired by a vibration sensor installed at a predetermined position of the target welding device; a calculation module, configured to determine a time-frequency energy distribution corresponding to the vibration data stream based on a collection time of the vibration data stream; a data processing module configured to perform signal filtering processing on the vibration data stream based on the energy proportion of the frequency band corresponding to the time-frequency energy distribution; and perform feature enhancement processing on the vibration data stream after filtering based on the mutation signal corresponding to the time-frequency energy distribution; an execution module, configured to generate data to be identified based on the enhanced vibration data stream, input the data to be identified into a preset deep residual neural network, perform fault category analysis on the vibration data stream, and determine a fault category analysis result; wherein the deep residual neural network is configured to extract rapidly changing features and stable structural features from the data to be identified using a preset dynamic gated attention mechanism, so as to analyze the fault category of the vibration data stream based on the rapidly changing features and the stable structural features; An output module, configured to determine a fault category corresponding to the target welding equipment according to the fault category analysis result; The data processing module is further configured to: calculate the frequency band energy corresponding to a preset time window based on the time-frequency energy distribution; determine the frequency band energy ratio corresponding to the time-frequency energy distribution based on the frequency band energy; determine the dominant frequency corresponding to the vibration data stream based on the frequency band energy ratio; and perform filtering processing on the preset frequency band of the vibration data stream based on the dominant frequency; The deep residual neural network includes multiple layers of time-frequency convolution modules, and the bandwidth of the time-frequency convolution modules at each layer is different; the execution module is further used to: input the data to be identified into a preset deep residual neural network, and extract the target local features of the data to be identified at different time scales and frequency ranges through the multiple layers of time-frequency convolution modules of the deep residual neural network; perform feature fusion on the target local features through a preset dynamic gated attention mechanism to generate a feature fusion vector; map the feature fusion vector to the fault category space to determine the fault category analysis result corresponding to the data to be identified; The execution module is also used to: extract the transient features of the target local features at each level through a convolution operation, and determine the steady-state features of the target local features at each level through a pooling method; through a preset gating unit, weightedly splice the transient features and the steady-state features of multiple levels according to the spectral energy of the corresponding levels to generate a feature fusion vector mixed feature corresponding to the target local features.

Images