A pattern recognition method based on signal image conversion and large model inference

Abstract

The application relates to the technical field of pattern recognition, and discloses a pattern recognition method based on signal image conversion and large model reasoning, which comprises the following steps: acquiring an original one-dimensional time sequence signal segment and calculating a polar coordinate phase angle vector, mapping the polar coordinate phase angle vector into a two-dimensional topological feature image by using a Gram angle field transformation; cutting a multi-scale feature image into image blocks, determining a dynamic shielding threshold according to global pixel variance, and mapping image blocks with local two-dimensional gradient variance lower than the threshold into background visual word units, and mapping the remaining image blocks into high-frequency mutation visual word units; splicing the visual word units and combining working condition constraint texts to construct a context sequence, inputting the context sequence into a reasoning model to calculate cross attention weights, and outputting classification results and causal deduction chains; and the application uses asymmetric resolution mapping and cross-modal attention focusing, eliminates background interference, and improves mutation feature recognition precision and deduction transparency.

Description

Technical Field

[0001] This invention relates to a pattern recognition method based on signal-to-image conversion and large-model inference, belonging to the field of pattern recognition technology. Background Technology

[0002] Utilizing deep neural networks to process one-dimensional time series signals has become a common technique in pattern recognition. Existing methods typically convert one-dimensional sequences into two-dimensional topological feature images to extract nonlinear global correlations between different time scales. Through this conversion, the nonlinear variations in one-dimensional time series signals are represented as spatial textures, while preserving the topological correlations between different frequency bands. As monitoring systems increasingly demand the capture of early, weak transient anomalies, the signal acquisition process needs to maintain a high sampling rate, which generates a large amount of redundant data reflecting low-frequency steady-state processes. For long-period topological images formed by high-frequency sampling, existing systems are limited by computational power and storage space, and typically use images of a uniform scale. Block aggregation or uniform dimensionality reduction are used, but steady-state operating data accounts for a large proportion of long-period signals. The smooth background texture formed by this data can distract the model's attention. During unified pooling or dimensionality reduction, high-frequency transient textures generated by local abrupt changes are easily weakened or even covered by the smooth background, resulting in the loss of a small proportion of abnormal features. To reduce background data, some methods use front-end amplitude thresholds to directly truncate part of the signal. However, this process destroys the temporal continuity and complete spatiotemporal topology of the original signal. Other methods expand the attention window of the basic inference model to enable it to receive full-resolution images. However, the computational cost of cross-attention increases rapidly with the input size, making it difficult to meet the computational cost requirements of practical applications.

[0003] Besides the aforementioned limitations in hardware physical processing such as data sampling and front-end filtering, existing technologies also fail to solve the problem of adaptive matching of background redundancy and local abrupt feature extraction mechanisms in subsequent software control algorithms. For example, Chinese invention patent CN118731689B discloses a preventive maintenance diagnosis method for motor equipment based on multi-parameter verification. Although this scheme maps one-dimensional signals to two-dimensional Gram angle field images, it uses preset fixed Gabor functions and geometric functions to perform global sliding window convolution on the image. This static feature extraction mechanism ignores the dynamic differences in pixel gradients between different regions of the image, cannot adaptively compress and smooth background redundancy, and causes high-frequency transient impact textures to be diluted by background information during global feature tiling. In addition, this scheme only outputs a single classification result based on fixed rule mapping, lacking deep integration of cross-modal working condition information and causal logic inference, making it difficult to meet the dual requirements of recognition accuracy and inference transparency for complex working conditions.

[0004] Therefore, the technical problem to be solved by this invention is how to use different resolution feature processing methods for different regions based on the differences in topological texture formed by the underlying signal, so as to retain high-frequency transient details while compressing redundant background data, and concentrate limited computing resources on key abnormal regions. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of the present invention is as follows: A pattern recognition method based on signal-to-image conversion and large-model inference, comprising the following steps:

[0006] Step S1: Obtain the original one-dimensional time series signal of the target object and extract the signal segment, calculate the phase angle vector of the signal segment in the polar coordinate system, and map it into a two-dimensional topological feature image using the Gram angle field transformation mechanism;

[0007] Step S2: Input the two-dimensional topological feature image into the visual encoder to extract multi-scale spatial feature maps, and use a linear projection layer to divide it into non-overlapping image blocks;

[0008] Step S3: Calculate the local two-dimensional gradient variance and global pixel variance of each image block, determine the dynamic occlusion threshold based on the global pixel variance, perform average pooling downsampling on image blocks with local two-dimensional gradient variance less than the dynamic occlusion threshold and map them as background visual words, and map image blocks with local two-dimensional gradient variance not less than the dynamic occlusion threshold as high-frequency mutation visual words according to the original feature resolution.

[0009] Step S4: Obtain the current working condition constraint text of the target object and convert it into a text word sequence. Concatenate background visual words and high-frequency mutation visual words according to the sequence dimension to form a continuous visual feature word sequence with non-uniform resolution distribution characteristics. Concatenate the text word sequence and the continuous visual feature word sequence to generate a multimodal context input sequence. Input the sequence into a large-scale basic inference model. Calculate the cross-attention weights between high-frequency mutation visual words, background visual words, and text words using a self-attention mechanism. Infer the spatiotemporal pattern evolution law of the two-dimensional topological feature image. Output a structured judgment text containing pattern classification results and causal logical inference chain.

[0010] Preferably, the calculation of the local two-dimensional gradient variance and global pixel variance of each image block in step S3, and the determination of the dynamic occlusion threshold based on the global pixel variance, includes the following sub-steps: Step S31, calculate the first-order spatial difference of each pixel in the horizontal and vertical directions within each image block to obtain the local two-dimensional spatial pixel gradient of each pixel; Step S32, based on the local two-dimensional spatial pixel gradient of each pixel, calculate the discrete variance of all pixel gradient values ​​within the corresponding image block, and use it as the local two-dimensional gradient variance of the image block; Step S33, summarize the gradient values ​​of all pixels in the multi-scale spatial feature map, calculate the mean and overall variance of the spatial pixel gradient of the entire image, and determine the dynamic occlusion threshold by multiplying the overall variance by the preset modulation coefficient; Step S34, compare the local two-dimensional gradient variance of each image block with the dynamic occlusion threshold in real time. When the local two-dimensional gradient variance is lower than the dynamic occlusion threshold, the corresponding image block is determined to belong to the low-frequency steady-state background. When the local two-dimensional gradient variance is not less than the dynamic occlusion threshold, the corresponding image block is determined to belong to the high-frequency transient impact region.

[0011] Preferably, the process of downsampling and mapping image blocks with local two-dimensional gradient variance less than the dynamic occlusion threshold to background visual words, and mapping image blocks with local two-dimensional gradient variance not less than the dynamic occlusion threshold to high-frequency mutation visual words according to the original feature resolution, includes the following sub-steps: Step S35, for image blocks determined to belong to low-frequency steady-state background, a two-dimensional spatial average pooling kernel is used to downsample the feature matrix of the image block, compressing the multi-dimensional pixel features into a single low-resolution feature vector, and generating a single background visual word through linear transformation; Step S36, for image blocks determined to belong to high-frequency transient impact regions, the original pixel feature resolution of the image block is maintained, and multiple high-resolution high-frequency mutation visual words are generated through full mapping by a linear projection layer; Step S37, according to the spatial topological array order of the original image blocks, the background visual words and each high-frequency mutation visual word are interwoven and spliced ​​to form a continuous visual feature word sequence with spatial non-uniform resolution distribution.

[0012] Preferably, the step S4 of obtaining the current operating condition constraint text of the target object and converting it into a text word sequence, and concatenating it with the continuous visual feature word sequence, includes the following sub-steps: Step S41, based on a preset structured prompt word template, obtain the target object's equipment operation log, safe service cycle stage, historical fault mode library, and current operating condition parameters, and combine them to construct the current operating condition constraint text; Step S42, segment and semantically encode the current operating condition constraint text using a preset word segmenter, and map it to generate a text word sequence; Step S43, use the text word sequence as a pre-guiding identifier, and concatenate it with the continuous visual feature word sequence in an autoregressive sequence order to generate a multimodal context input sequence containing domain expert prior constraints and image topological details.

[0013] Preferably, the calculation of cross-attention weights between high-frequency mutation visual words, background visual words, and text words using the self-attention mechanism includes: in the self-attention calculation layer of the large-scale basic inference model, shielding the self-attention calculation weight allocation between background visual words, and concentrating the self-attention calculation resources on the cross-attention correlation calculation between text words and high-frequency mutation visual words; based on the cross-attention correlation calculation results, establishing an isomorphic mapping relationship between the semantic vector representing the specific fault evolution trend in the current working condition constraint text and the feature vector representing the violent phase angle oscillation texture in the high-frequency mutation visual words.

[0014] Preferably, before obtaining the original one-dimensional time series signal of the target object in step S1, the method further includes: setting the discretization sampling rate to 10kHz according to the Nyquist sampling theorem; normalizing the amplitude of the extracted signal segment and linearly mapping the amplitude to the interval between -1 and 1; setting the time-domain sliding window slice length of the extracted signal segment to 1024 sampling points; and setting the size of each image block obtained by the linear projection layer when segmenting the multi-scale spatial feature map to 16×16 pixels.

[0015] Preferably, the structured judgment text outputting the pattern classification result and the causal logic inference chain includes: outputting a diagnostic classification label representing the current operating state of the target object; outputting the corresponding causal logic inference chain, which includes a location description of the high-frequency texture distortion position in the two-dimensional topological feature image, an anomaly confidence obtained by combining the current operating condition constraint text, and a technical causal explanation text for determining the current operating state.

[0016] Preferably, after normalizing the amplitude of the intercepted signal segment, the method further includes: real-time monitoring of the signal variance volatility of the original one-dimensional time series signal; when the signal variance volatility exceeds a preset volatility threshold, automatically increasing the overlap rate of the time-domain sliding window of the intercepted signal segment from a preset initial overlap rate to a target overlap rate, so as to increase the mapping frequency of the transient impact region in the adjacent time-domain sliding window segment.

[0017] Preferably, after outputting the structured judgment text containing the pattern classification result and the causal logical deduction chain, the method further includes: when the pattern classification result indicates that the target object is in an abnormal fault state, generating a control protection command based on the abnormal confidence level and technical causal explanation text output in the structured judgment text; transmitting the control protection command to the control terminal of the target object, driving the control terminal to adjust the operating power of the target object, and triggering a shutdown response when the abnormal confidence level reaches a preset alarm threshold.

[0018] Compared with the prior art, the beneficial effects of the present invention are:

[0019] 1. In pattern recognition for signal image conversion and large model inference, one-dimensional time series signals are converted into two-dimensional topological feature images. Based on the local two-dimensional gradient variance of image blocks and the dynamic occlusion threshold, low-frequency steady-state background and high-frequency transient impact regions are distinguished. This can separate the background texture formed by stable operation from the abnormal texture formed by drastic phase angle changes, avoid the weak local abnormal features being masked by the background signal in long-term sampling data, and improve the distinguishability of signal features under complex working conditions.

[0020] 2. This invention performs average pooling downsampling on image blocks corresponding to low-frequency steady-state backgrounds, while preserving the original feature resolution of high-frequency transient impact regions. This not only compresses a large amount of repetitive background data, but also completely preserves the transient details of key regions, reducing the number of visual terms input to large-scale basic inference models, allowing limited computing resources to be used more for analyzing abnormal textures, and reducing attentional distraction caused by background information.

[0021] 3. This invention inputs the current operating condition constraint text and the continuous visual feature word sequence into a large-scale basic reasoning model, so that the high-frequency mutation features in the equipment operation information and the image can correspond to each other during the reasoning process. With the guidance of the operating condition information, the model no longer relies solely on numerical features for isolated classification, but can analyze abnormal textures in combination with actual operating conditions. In addition to outputting the pattern classification results, it can also provide the corresponding causal logic deduction chain, improving the transparency and interpretability of the abnormal state determination process. Attached Figure Description

[0022] Figure 1 This is a flowchart of the signal topology mapping and multimodal reasoning process of this invention;

[0023] Figure 2 This is a structural diagram of the edge computing pattern recognition system of the present invention.

[0024] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0026] A pattern recognition method based on signal-to-image conversion and large-model inference includes the following steps:

[0027] Step S1: Obtain the original one-dimensional time series signal of the target object and extract the signal segment, calculate the phase angle vector of the signal segment in the polar coordinate system, and map it into a two-dimensional topological feature image using the Gram angle field transformation mechanism;

[0028] Step S2: Input the two-dimensional topological feature image into the visual encoder to extract multi-scale spatial feature maps, and use a linear projection layer to divide it into non-overlapping image blocks;

[0029] Step S3: Calculate the local two-dimensional gradient variance and global pixel variance of each image block, determine the dynamic occlusion threshold based on the global pixel variance, perform average pooling downsampling on image blocks with local two-dimensional gradient variance less than the dynamic occlusion threshold and map them as background visual words, and map image blocks with local two-dimensional gradient variance not less than the dynamic occlusion threshold as high-frequency mutation visual words according to the original feature resolution.

[0030] Step S4: Obtain the current working condition constraint text of the target object and convert it into a text word sequence. Concatenate background visual words and high-frequency mutation visual words according to the sequence dimension to form a continuous visual feature word sequence with non-uniform resolution distribution. Concatenate the text word sequence and the continuous visual feature word sequence to generate a multimodal context input sequence. Input the sequence into a large-scale basic inference model. Calculate the cross-attention weights between high-frequency mutation visual words, background visual words, and text words using a self-attention mechanism. Infer the spatiotemporal pattern evolution law of the two-dimensional topological feature image. Output a structured judgment text containing pattern classification results and causal logical inference chain.

[0031] Preferably, the calculation of the local two-dimensional gradient variance and global pixel variance of each image block in step S3, and the determination of the dynamic occlusion threshold based on the global pixel variance, includes the following sub-steps: Step S31, calculate the first-order spatial difference of each pixel in the horizontal and vertical directions within each image block to obtain the local two-dimensional spatial pixel gradient of each pixel; Step S32, based on the local two-dimensional spatial pixel gradient of each pixel, calculate the discrete variance of all pixel gradient values ​​within the corresponding image block, and use it as the local two-dimensional gradient variance of the image block; Step S33, summarize the gradient values ​​of all pixels in the multi-scale spatial feature map, calculate the mean and overall variance of the spatial pixel gradient of the entire image, and determine the dynamic occlusion threshold by multiplying the overall variance by the preset modulation coefficient; Step S34, compare the local two-dimensional gradient variance of each image block with the dynamic occlusion threshold in real time. When the local two-dimensional gradient variance is lower than the dynamic occlusion threshold, the corresponding image block is determined to belong to the low-frequency steady-state background. When the local two-dimensional gradient variance is not less than the dynamic occlusion threshold, the corresponding image block is determined to belong to the high-frequency transient impact region.

[0032] Preferably, the process of downsampling and mapping image blocks with local two-dimensional gradient variance less than the dynamic occlusion threshold to background visual words, and mapping image blocks with local two-dimensional gradient variance not less than the dynamic occlusion threshold to high-frequency mutation visual words according to the original feature resolution, includes the following sub-steps: Step S35, for image blocks determined to belong to low-frequency steady-state background, a two-dimensional spatial average pooling kernel is used to downsample the feature matrix of the image block, compressing the multi-dimensional pixel features into a single low-resolution feature vector, and generating a single background visual word through linear transformation; Step S36, for image blocks determined to belong to high-frequency transient impact regions, the original pixel feature resolution of the image block is maintained, and multiple high-resolution high-frequency mutation visual words are generated through full mapping by a linear projection layer; Step S37, according to the spatial topological array order of the original image blocks, the background visual words and each high-frequency mutation visual word are interwoven and spliced ​​to form a continuous visual feature word sequence with spatial non-uniform resolution distribution.

[0033] Preferably, the step S4 of obtaining the current operating condition constraint text of the target object and converting it into a text word sequence, and concatenating it with the continuous visual feature word sequence, includes the following sub-steps: Step S41, based on a preset structured prompt word template, obtain the target object's equipment operation log, safe service cycle stage, historical fault mode library, and current operating condition parameters, and combine them to construct the current operating condition constraint text; Step S42, segment and semantically encode the current operating condition constraint text using a preset word segmenter, and map it to generate a text word sequence; Step S43, use the text word sequence as a pre-guiding identifier, and concatenate it with the continuous visual feature word sequence in an autoregressive sequence order connected head to tail, to generate a multimodal context input sequence containing domain expert prior constraints and image topological details.

[0034] Preferably, the calculation of cross-attention weights between high-frequency mutation visual words, background visual words, and text words using the self-attention mechanism includes: in the self-attention calculation layer of the large-scale basic inference model, shielding the self-attention calculation weight allocation between background visual words, and concentrating the self-attention calculation resources on the cross-attention correlation calculation between text words and high-frequency mutation visual words; based on the cross-attention correlation calculation results, establishing an isomorphic mapping relationship between the semantic vector representing the specific fault evolution trend in the current working condition constraint text and the feature vector representing the violent phase angle oscillation texture in the high-frequency mutation visual words.

[0035] Preferably, before obtaining the original one-dimensional time series signal of the target object in step S1, the method further includes: setting the discretization sampling rate to 10kHz according to the Nyquist sampling theorem; normalizing the amplitude of the extracted signal segment and linearly mapping the amplitude to the interval between -1 and 1; setting the time-domain sliding window slice length of the extracted signal segment to 1024 sampling points; and setting the size of each image block obtained by the linear projection layer when segmenting the multi-scale spatial feature map to 16×16 pixels.

[0036] Preferably, the structured judgment text outputting the pattern classification result and the causal logic inference chain includes: outputting a diagnostic classification label representing the current operating state of the target object; outputting the corresponding causal logic inference chain, which includes a location description of the high-frequency texture distortion position in the two-dimensional topological feature image, an anomaly confidence obtained by combining the current operating condition constraint text, and a technical causal explanation text for determining the current operating state.

[0037] Preferably, after normalizing the amplitude of the intercepted signal segment, the method further includes: real-time monitoring of the signal variance volatility of the original one-dimensional time series signal; when the signal variance volatility exceeds a preset volatility threshold, automatically increasing the overlap rate of the time-domain sliding window of the intercepted signal segment from a preset initial overlap rate to a target overlap rate, so as to increase the mapping frequency of the transient impact region in the adjacent time-domain sliding window segment.

[0038] Preferably, after outputting the structured judgment text containing the pattern classification result and the causal logical deduction chain, the method further includes: when the pattern classification result indicates that the target object is in an abnormal fault state, generating a control protection command based on the abnormal confidence level and technical causal explanation text output in the structured judgment text; transmitting the control protection command to the control terminal of the target object, driving the control terminal to adjust the operating power of the target object, and triggering a shutdown response when the abnormal confidence level reaches a preset alarm threshold.

[0039] Example 1: The current data acquisition terminal monitors the non-stationary vibration of a high-speed train axle box bearing during its high-speed operation. Based on the Nyquist sampling theorem, the original one-dimensional time series signal of the bearing operation is acquired at a sampling rate of 10kHz. A time-domain sliding window with a length of 1024 sampling points is used to extract a segment of the signal. The original amplitude of the signal segment is linearly mapped to the interval between -1 and 1 to obtain a normalized signal. Then, the phase vector is calculated according to the polar coordinate system. ,in, Represents the normalized signal. This represents the corresponding phase angle vector, representing the time step. and time step Corresponding to 0 to Polar phase angle within the interval and Substituting into the Gram angle difference field mapping formula in the Gram angle field transformation mechanism Generate a two-dimensional topological feature image with a size of 1024×1024, where, and Indicates the time step index. and Representing time steps and time step The corresponding polar phase angle, The index position in the two-dimensional topological feature image is The elements in the image matrix represent the inner product correlation between signal amplitudes at different time steps. The stable operation cycle corresponds to a uniformly distributed low-frequency steady-state background, and the early peeling of the inner circle forms a local phase angle oscillation texture.

[0040] When the train runs at a speed of 300 km / h, the rotation frequency of the axle box bearing is about 45 Hz, and the single mechanical rotation cycle is 22 ms. The high-frequency structural self-vibration impact caused by the early spalling failure of the bearing inner ring usually lasts for 2 ms to 5 ms. A time-domain sliding window consisting of 1024 sampling points and a sampling rate of 10 kHz is used. The window duration is 0.1024 s, which can cover about 4.5 bearing mechanical rotation cycles and capture 4 to 5 high-frequency transient impact events within a single window. This window contains both the rotation cycle information required to form a low-frequency steady-state background and makes the high-frequency transient impact region form a recognizable phase angle oscillation texture in the diagonal neighborhood of the two-dimensional topological feature image, thereby matching the signal sampling scale with the bearing mechanical motion cycle.

[0041] The visual encoder receives a two-dimensional topological feature image and extracts feature matrices of different spatial scales layer by layer to form a multi-scale spatial feature map. The linear projection layer divides the multi-scale spatial feature map into a fixed grid of 16×16 pixels, so that each pixel is assigned to only one image block, resulting in a non-overlapping image block of size 16×16. For each image block, the first-order spatial difference of each pixel in the horizontal and vertical directions is calculated to obtain the local two-dimensional spatial pixel gradient. The discrete variance of the gradient values ​​of all pixels in the image block is calculated and used as the local two-dimensional gradient variance of the image block.

[0042] The gradient values ​​of all pixels in the multi-scale spatial feature map are summarized, the mean of the pixel gradients in the entire image space and the overall variance of each gradient value relative to the mean are calculated, and the overall variance is used as the global pixel variance. The global pixel variance is multiplied by a preset modulation coefficient to determine the discrete decision boundary value. The preset modulation coefficient is obtained by continuously acquiring 100 sets of baseline signal segments under fault-free and stable operating conditions, extracting the local gradient variance sequence of each 16×16 pixel image block and calculating the 99th percentile value. The 99th percentile value is divided by the overall variance of the baseline image to obtain a fixed dimensionless constant of 1.2. The result of multiplying the dimensionless constant 1.2 by the real-time overall variance of the image is determined as the dynamic occlusion threshold, so that the dynamic occlusion threshold is adjusted with the global gradient change of the current two-dimensional topological feature image.

[0043] The local two-dimensional gradient variance of each image block is compared with a dynamic occlusion threshold. Image blocks outside the diagonal are identified as low-frequency steady-state backgrounds when their local two-dimensional gradient variance is less than the dynamic occlusion threshold. A two-dimensional spatial average pooling kernel is used to downsample the feature matrix of these image blocks, compressing multi-dimensional pixel features into a single low-resolution feature vector. This vector is then linearly transformed to generate a single background visual term. Image blocks in the diagonal neighborhood are identified as high-frequency transient impact regions when their local two-dimensional gradient variance is not less than the dynamic occlusion threshold. These image blocks retain their original feature resolution and are fully mapped to multiple high-frequency abrupt visual terms through a linear projection layer. Following the spatial topological array order of the original image blocks, the... Background visual words and high-frequency abrupt visual words are interwoven and spliced ​​together. The single low-resolution background visual word generated by downsampling is expanded into a 512-dimensional feature vector by the first linear projection matrix. The 16 high-resolution high-frequency abrupt visual words that retain the original resolution are dimensionally reduced and mapped into 16 identical 512-dimensional feature vectors by the second linear projection matrix. Then, each 512-dimensional feature vector is element-wise added and bound to a 512-dimensional two-dimensional spatial position vector obtained by converting the row and column coordinates of the image block grid through a sinusoidal position encoder. This forms a continuous visual feature word sequence with a non-uniform resolution distribution, which compresses the low-frequency steady-state background while retaining local high-frequency transient impact details.

[0044] The system reads the target object's equipment operation log, safe service cycle stage, historical fault mode library, and current operating condition parameters according to a preset structured prompt word template, and combines them to form the current operating condition constraint text. In this embodiment, the equipment operation log records the cumulative service information of the axle box bearing, the safe service cycle stage indicates that the current bearing's service cycle has reached 80% of the safe threshold, the historical fault mode library provides the high-frequency transient impact mode corresponding to the initial stage of inner ring spalling, and the current operating condition parameters include train running speed and bearing vibration monitoring parameters. Based on the above information, the current operating condition constraint text sets the high-frequency transient impact characteristics of the initial stage of inner ring spalling as the key focus content.

[0045] The word segmenter segments and semantically encodes the current working condition constraint text to generate a text word sequence. The text word sequence is used as a pre-guided identifier and concatenated with the continuous visual feature word sequence in an autoregressive sequence order to generate a multimodal context input sequence. This sequence contains both domain expert prior constraints and spatial texture information from the two-dimensional topological feature image, and is then input into a large-scale basic inference model with 7 billion parameters.

[0046] In the self-attention computation layer of the large-scale basic inference model, a binarized attention mask matrix is ​​set to shield the self-attention weight allocation between background visual words and retain the cross-attention computation channel between text words and high-frequency mutation visual words. The model reads the continuous visual feature word sequence according to the spatial topological order of image blocks, compares the feature differences between high-frequency mutation visual words and adjacent background visual words, and combines the safe service cycle stage, historical failure mode and current operating condition parameters in the current working condition constraint text to deduce the spatiotemporal pattern evolution law between the high-frequency transient impact region and the low-frequency steady-state background in the two-dimensional topological feature image.

[0047] Based on the cross-attention correlation calculation results, the model establishes an isomorphic mapping relationship between the semantic vector representing the specific fault evolution trend in the current working condition constraint text and the feature vector representing the texture of violent phase angle oscillation in the high-frequency mutation visual words. It then outputs structured judgment text using an autoregressive generation method. This structured judgment text is generated using a preset JSON nested field format. Specifically, the causal logical deduction chain and the technical causal explanation text are represented by a logical reasoning structure consisting of three discrete key-value pairs: the first key-value pair is the texture distortion location field, composed of the row and column indices of the image matrix extracted based on the two-dimensional spatial coordinates of each high-frequency mutation visual word; the second key-value pair is calculated by taking the maximum value of the cross-attention correlation matrix between the text words and the high-frequency mutation visual words. The discrete floating-point number 0.92 is used as the anomaly confidence field; the third key-value pair is a fault attribution description string obtained by combining the location field with the prior rule in the operating condition constraint text that the bearing service life reaches the safety threshold of 80%. The structured judgment text includes an early fault diagnosis classification label for the inner ring, a location description of the high-frequency texture distortion position in the two-dimensional topological feature image, anomaly confidence, and technical causal explanation text. The anomaly confidence is calculated based on the cross-attention correlation between the current operating condition constraint text and the high-frequency mutation visual word. The causal logic inference chain records that dense high-frequency texture distortion appears in the diagonal region of the two-dimensional topological feature image. Combined with the constraint condition that the bearing service life reaches the safety threshold of 80%, the high-frequency texture distortion is judged to be consistent with the impact characteristics of the initial stage of inner ring spalling.

[0048] In the low-rank adaptive fine-tuning stage before model deployment, most of the weights of the pre-trained model are fixed, the original parameters of the cross-attention projection layer are frozen, only the additional low-rank matrix is ​​trained, and the model output is constrained by the contrastive learning loss function. The hidden layer feature vectors obtained by linear transformation of high-frequency mutation visual words are placed in the same metric space as the semantic word vectors representing the fault evolution trend in the current working condition constraint text. By minimizing the cosine distance between the two, the geometric distance alignment and directional consistency mapping of the cosine similarity of the two types of feature vectors in the 512-dimensional latent space are achieved, avoiding the misuse of abstract isomorphic terms, thereby corresponding the phase angle oscillation texture in the two-dimensional topological feature image with the fault evolution semantics in the current working condition constraint text.

[0049] After the above processing, the low-frequency steady-state background generated by long-period high-frequency sampling is downsampled and compressed during the visual feature word generation stage, while the high-frequency transient impact region is preserved according to the original feature resolution. After the self-attention weights between background visual words are masked, the cross-attention calculation is concentrated between text words and high-frequency abrupt visual words, avoiding the self-attention calculation amount from increasing with the square of the sequence length due to full-resolution visual words, and also reducing the influence of low-frequency steady-state background on local high-frequency texture. After the current operating condition constrained text participates in the cross-attention calculation, the large-scale basic inference model can interpret the local high-frequency texture based on the equipment service stage, historical failure modes and current operating condition parameters, and output the pattern classification results, anomaly confidence and causal logic inference chain. Thus, a pattern recognition process is formed from the original one-dimensional time series signal, two-dimensional topological feature image, multi-scale spatial feature map and continuous visual feature word sequence to structured judgment text.

[0050] Example 2: The current accelerometer continuously collects vibration acceleration signals during the high-speed operation of the axle box bearing of an electric locomotive. Its measurement range is 0.1Hz to 20kHz, the full scale is ±50g, and the sampling rate is 10kHz. To simulate electromagnetic interference and mechanical background noise in industrial environments, Gaussian white noise with a signal-to-noise ratio of 20dB and 50Hz power frequency interference harmonics are superimposed on the collected signal to form an original one-dimensional time series signal. A time-domain sliding window with a length of 1024 sampling points is used to extract the signal segment, and its amplitude is normalized to the interval between -1 and 1. The phase angle vector is calculated according to the inverse cosine function, and then a two-dimensional topological feature image with a size of 1024×1024 is generated through the Gram angle field transformation mechanism.

[0051] The sampling rate and the length of the time-domain sliding window are determined based on the response bandwidth of the high-frequency transient impact and the computational complexity of the two-dimensional topological feature image. When the sampling rate is lower than 5kHz, aliasing and frame dropping occur in the high-frequency phase angle abrupt texture. When the length of the time-domain sliding window is greater than 2048 sampling points, the size of the generated image matrix increases, and the real-time computational complexity exceeds the processing capacity of the experimental platform. Therefore, this embodiment uses a 10kHz sampling rate and a time-domain sliding window of 1024 sampling points to conduct subsequent experiments.

[0052] The experimental platform uses an edge computing node equipped with a single graphics card, with a video memory capacity of 24GB and a video memory bandwidth of 1008GB / s. The inference model is loaded with half-precision floating-point operations. The test dataset contains 10,000 vibration acceleration samples, including 5,000 samples of normal operation, 2,500 samples of inner ring peeling, and 2,500 samples of outer ring peeling. The pattern recognition accuracy is obtained by comparing the predicted classification results with the real physical labels. The model inference latency is obtained by statistically analyzing the average end-to-end time of 10,000 consecutive inferences.

[0053] To compare the impact of each technical step on visual word compression, attention allocation, and pattern recognition results, the experiment set up the present invention sample group, a partially missing control group A, a partially missing control group B, a traditional temporal control group C, an out-of-range control group D, and an out-of-range control group E. In the present invention sample group, a two-dimensional topological feature image is input into a visual encoder to extract multi-scale spatial feature maps. The image is divided into non-overlapping image blocks by a linear projection layer. The local two-dimensional gradient variance and global pixel variance of each image block are calculated. A dynamic occlusion threshold is used to distinguish between low-frequency steady-state background and high-frequency transient impact regions. For low-frequency steady-state background, the 256-dimensional pixel features corresponding to the 16×16 feature matrix are averaged and downsampled into a single background visual word. For high-frequency transient impact regions, the original feature resolution is maintained and high-frequency abrupt visual words are generated. Subsequently, the continuous visual feature word sequence is concatenated with the text word sequence to form a multimodal context input sequence, which is input into a large-scale basic inference model for cross-attention calculation and pattern classification.

[0054] The partially missing control group A removed the comparison process between the local two-dimensional gradient variance and the dynamic occlusion threshold, and applied global uniform pooling to all image blocks to generate equal-length visual word sequences. The partially missing control group B did not perform background visual word downsampling, and all image blocks were mapped according to the original feature resolution, generating a total of 4096 visual words. The traditional time series control group C did not perform signal-to-image conversion, and used a combination of 1D-CNN and LSTM to process the original one-dimensional time series signal. The out-of-range control group D set the preset modulation coefficient to 0.2, so that many low-frequency steady-state backgrounds were not averaged. The out-of-range control group E set the preset modulation coefficient to 2.5, so that some high-frequency transient impact regions were identified as low-frequency steady-state backgrounds and averaged.

[0055] Under stable operating conditions, the sampling signal signal-to-noise ratio of the sample group of this invention is 20dB, the root mean square background noise is 0.42m / s², and after comparing the local two-dimensional gradient variance with the dynamic occlusion threshold, the continuous visual feature word sequence contains 388 visual words, the attention weight of key high-frequency regions accounts for 88.4%, the model inference latency is 14.2ms, and the pattern recognition accuracy is 98.6%.

[0056] Under early fault conditions in the inner circle, the sampling signal signal-to-noise ratio of the sample group of this invention is 20dB, the root mean square background noise is 0.45m / s², after the local high-frequency transient impact area is preserved according to the original feature resolution, the continuous visual feature word sequence contains 412 visual words, the attention weight ratio of key high-frequency areas is 92.1%, the model inference delay is 14.8ms, and the pattern recognition accuracy is 98.2%.

[0057] Under the outer ring peeling fault condition, the sampling signal signal-to-noise ratio of the sample group of this invention is 20dB, the root mean square background noise is 0.43m / s², and after non-uniform resolution mapping, the continuous visual feature word sequence contains 405 visual words, the key high-frequency region attention weight accounts for 91.5%, the model inference latency is 14.5ms, and the pattern recognition accuracy is 98.4%. After compressing 4096 full-resolution visual words to 388 to 412 visual words, the word compression rate exceeds 89%. Under both fault conditions, the key high-frequency region attention weight accounts for more than 91%, the model inference latency under all three conditions is controlled within 14.8ms, and the pattern recognition accuracy remains above 98%.

[0058] The signal-to-noise ratio of the sampled signal in the partially missing control group A was 20dB, the root mean square background noise was 0.42m / s², and 400 visual words were generated after global uniform pooling. The model inference latency was 14.3ms. Due to the pooling of local phase angle abrupt texture and low-frequency steady-state background, the attention weight ratio of key high-frequency regions was reduced to 18.2%, and the pattern recognition accuracy was 71.4%.

[0059] The signal-to-noise ratio of the sampled signal in the partially missing control group B was 20dB, and the root mean square background noise was 0.45m / s². After all visual features were input into the large-scale basic inference model at the original feature resolution, the attention weight of key high-frequency regions accounted for 32.5%, and the pattern recognition accuracy was 89.1%. The addition of 4096 visual words increased the computational cost of self-attention, resulting in a model inference latency of 215.6ms.

[0060] The traditional time series control group C had a sampling signal-to-noise ratio of 20dB and a root mean square background noise of 0.42m / s². This group directly used 1D-CNN and LSTM to process the original one-dimensional time series signal without generating visual words. Therefore, the number of generated visual words and the proportion of attention weights in key high-frequency regions were not counted. Its model inference latency was 8.5ms and the pattern recognition accuracy was 62.8%.

[0061] The sampling signal signal-to-noise ratio of the out-of-range control group D was 20dB, the root mean square background noise was 0.43m / s², the preset modulation coefficient was set to 0.2, the dynamic occlusion threshold was low, a lot of low-frequency steady-state background was retained according to the original feature resolution, the continuous visual feature word sequence contained 2850 visual words, the attention weight of key high-frequency regions accounted for 45.6%, the model inference latency was 125.4ms, and the pattern recognition accuracy was 82.3%.

[0062] The sampling signal signal-to-noise ratio of the out-of-range control group E was 20dB, the root mean square background noise was 0.45m / s², the preset modulation coefficient was set to 2.5, the dynamic occlusion threshold was too high, some high-frequency transient impact regions were averaged, the continuous visual feature word sequence contained only 256 visual words, the attention weight of key high-frequency regions accounted for 12.1%, the model inference latency was 11.2ms, and the pattern recognition accuracy was 68.2%.

[0063] The above data shows that although the partially missing control group A has a similar number of visual words and model inference latency as the sample group of this invention, global uniform pooling weakens local high-frequency textures, making it difficult to focus attention calculations on fault regions. The partially missing control group B retains complete visual features, but as the number of visual words increases, the self-attention calculation workload increases with the sequence length, and the inference latency increases accordingly. The out-of-range control group D retains too much low-frequency steady-state background, while the out-of-range control group E compresses some high-frequency transient impact regions. The pattern recognition accuracy of both groups is lower than that of the sample group of this invention. It can be seen that by using the local two-dimensional gradient variance and global pixel variance to determine the dynamic occlusion threshold and processing the low-frequency steady-state background and high-frequency transient impact regions respectively, it is possible to retain fault textures while compressing background visual words, so that cross-attention calculations are focused on high-frequency abrupt visual words and text words.

[0064] Example 3: The current signal processing unit monitors the non-stationary vibration of the axle box bearing during high-load operation of the high-speed train. It acquires the original one-dimensional time series signal at a sampling rate of 10kHz, uses a time-domain sliding window with a length of 1024 sampling points to extract the signal segment, normalizes the signal amplitude to the interval of -1 to 1, calculates the phase vector according to the inverse cosine function, and then uses the Gram angle field transform mechanism to generate a two-dimensional topological feature image with a size of 1024×1024.

[0065] The visual encoder receives a two-dimensional topological feature image and extracts a multi-scale spatial feature map. The linear projection layer then divides the multi-scale spatial feature map into non-overlapping image blocks of size 16×16. The image is divided into blocks, denoted as... For the coordinates within this image block, pixels, with The pixel feature values ​​at the corresponding locations are represented by the first-order differences in the horizontal and vertical spatial directions, respectively: ,in, Represents pixels First-order spatial difference in the horizontal direction This represents the first-order spatial difference of the pixel in the vertical direction. The local two-dimensional spatial pixel gradient of each pixel is calculated based on the first-order spatial difference between the two directions, and the image is divided into blocks. The discrete variance of the gradient values ​​of all pixels within the image block is used as the local two-dimensional gradient variance of that image block. .

[0066] Summarize the gradient values ​​of all pixels in the multi-scale spatial feature map, calculate the mean and overall variance of the pixel gradients across the entire image, and denot the overall variance as the global pixel variance. The preset modulation coefficient is denoted as The dynamic occlusion threshold is denoted as Both conditions are met: In this embodiment, the preset modulation coefficient is... The value is set to 1.2. This value is determined based on the ratio between the maximum gradient variance envelope of the background clutter measured under the stable baseline condition and the global pixel variance. Since the global pixel variance changes with the image gradient under the current condition, the dynamic occlusion threshold is also adjusted accordingly.

[0067] The local two-dimensional gradient variance of each image block is compared with the dynamic occlusion threshold. Less than At that time, the image is divided into blocks. Determined as a low-frequency steady-state background, a two-dimensional spatial average pooling kernel is used to downsample its 16×16 feature matrix, compressing the 256-dimensional pixel feature vector into a single low-resolution feature vector. This is then transformed linearly to generate a single background visual word. Not less than At that time, the image is divided into blocks and identified as high-frequency transient impact regions. Average pooling is not performed to maintain its original feature resolution. The original pixel features are fully mapped by the linear projection layer to generate 16 high-frequency mutation visual terms.

[0068] Following the spatial topological array order of the original image blocks, background visual words and high-frequency abrupt visual words are interwoven and spliced ​​to form a continuous sequence of visual feature words. Before stitching, two-dimensional absolute spatial coordinates corresponding to the row and column indices of the original image array grid are set for each image block. For an image block compressed into a single background visual word, the two-dimensional coordinates of the center grid of the image block are used as the position identifier. For an image block that generates 16 high-frequency mutation visual words, each high-frequency mutation visual word is bound to its corresponding subdivided grid two-dimensional coordinates. The two-dimensional sinusoidal position encoder converts each two-dimensional coordinate into a position vector and adds the position vector to the corresponding visual word feature vector, so that each word in the continuous visual feature word sequence retains its spatial position in the two-dimensional topological feature image.

[0069] The word segmenter converts the current operating condition constraint text, which includes equipment service life, key monitoring locations, and current operating condition parameters, into a text word sequence. Then, the text word sequence is used as a pre-guiding identifier and concatenated with the continuous visual feature word sequence in an autoregressive order to generate a multimodal context input sequence. This is a large-scale basic inference model with a sequence of input parameters of size 7B.

[0070] An attention mask matrix is ​​set in the self-attention computation layer of a large-scale basic inference model. The self-attention weight allocation between background visual words is masked, and computational resources are concentrated on calculating the cross-attention correlation between text words and high-frequency mutation visual words. Based on the cross-attention correlation, the large-scale basic inference model establishes an isomorphic mapping relationship between the semantic vector representing the fault evolution trend in the current working condition constraint text and the feature vector representing the phase angle violent oscillation texture in the high-frequency mutation visual words. It outputs a structured judgment text containing pattern classification labels, anomaly confidence and causal logic inference chain. The anomaly confidence is generated based on the cross-attention correlation between text words and high-frequency mutation visual words and the pattern classification result.

[0071] During bearing operation, the signal processing unit uses 100 consecutive time-domain sliding windows as a monitoring cycle, calculates the average variance of each signal segment in the current monitoring cycle and the previous monitoring cycle, and uses the change ratio of the two average values ​​to characterize the signal variance fluctuation rate. When the signal variance fluctuation rate exceeds the preset fluctuation threshold of 15%, the overlap rate of the time-domain sliding window is increased from the preset initial overlap rate of 50% to the target overlap rate of 75%, thereby increasing the mapping frequency of the high-frequency transient impact region in adjacent time-domain sliding window segments.

[0072] The large-scale basic inference model generates structured judgment text based on the cross-attention calculation results, locates the high-frequency phase angle oscillation position in the diagonal region of the two-dimensional topological feature image, and combines the information that the bearing service cycle exceeds 80% in the current working condition constraint text to output the pattern classification result of the initial failure of inner ring peeling. The causal logic inference chain records the high-frequency texture distortion position, the service cycle condition in the current working condition constraint text, and the correlation between the two.

[0073] After the structured judgment text is output, the control conversion module uses a regular expression parser to extract diagnostic classification labels and anomaly confidence levels. When the diagnostic classification label indicates that the bearing is in an abnormal fault state and the anomaly confidence level reaches the preset alarm threshold of 85%, the control conversion module generates a control protection command and encapsulates it into a standard command frame according to the industrial control protocol. In this frame, the power adjustment parameters are written to the control register address, and the emergency stop flag is written to the switch output register. The encapsulated control protection command is transmitted to the programmable logic controller terminal of the target object through the Ethernet interface, driving the control terminal to reduce the operating power at a set rate, or to perform an emergency stop when the anomaly confidence level reaches the preset alarm threshold.

[0074] Example 4: This example combines Figures 1 to 2 This paper describes a pattern recognition method based on signal-to-image conversion and large-model inference, such as... Figure 1 As shown, step S1 obtains the original one-dimensional time series signal of the target object and extracts signal segments, calculates the phase vector of the signal segments in the polar coordinate system, and maps it to a two-dimensional topological feature image using Gram angle field transformation; step S2 inputs the two-dimensional topological feature image into a visual encoder to extract multi-scale spatial feature maps, and divides it into non-overlapping image blocks using a linear projection layer; step S3 calculates the local two-dimensional gradient variance and global pixel variance of each image block, determines the dynamic occlusion threshold based on the global pixel variance, performs average pooling downsampling on image blocks with local two-dimensional gradient variance less than the dynamic occlusion threshold to map them to background visual words, and maps image blocks with local two-dimensional gradient variance not less than the dynamic occlusion threshold to high-frequency mutation visual words according to the original feature resolution; step S4 obtains the current working condition constraint text of the target object and converts it into a text word sequence, splices the background visual words and high-frequency mutation visual words according to the sequence dimension to generate a multimodal context input sequence, inputs it into a large-scale basic inference model, uses a self-attention mechanism to infer the spatiotemporal pattern evolution law of the two-dimensional topological feature image, and outputs a structured judgment text containing pattern classification results and causal logic inference chain.

[0075] like Figure 2 As shown, the data acquisition terminal uses an accelerometer to collect and transmit the original one-dimensional time series signal to the edge computing node. The equipment ledger management terminal reads the data in the historical operation database to generate the current operating condition constraint text and transmits it to the edge computing node. The edge computing node is equipped with a visual encoder, a word segmenter, a large-scale basic inference model, and a control conversion module. The visual encoder and word segmenter process the input signal features and the current operating condition constraint text, respectively, and provide them to the large-scale basic inference model. The control conversion module generates control protection commands based on the processing results and transmits them to the programmable logic controller terminal using the Ethernet interface industrial control protocol.

[0076] Example 5: To address channel noise introduced by different sensing interfaces and signal transmission links, the signal processing unit performs preset modulation coefficient calibration before bearing monitoring. The bearing is kept running smoothly without load. The signal processing unit continuously collects 100 sets of fault-free vibration acceleration signal segments. Each signal segment contains 1024 sampling points with a sampling rate of 10kHz. Each signal segment is converted into a two-dimensional topological feature image. The spatial gradient value of all pixels in each image and the corresponding global pixel variance are calculated. The mean of the global pixel variance of the 100 images is then taken as the global pixel variance benchmark under fault-free background.

[0077] In the calibration data mentioned above, the local two-dimensional gradient variance of each image block is statistically analyzed, and the maximum variance envelope value corresponding to the fault-free background clutter is obtained. The ratio of this maximum variance envelope value to the global pixel variance benchmark is used as the preset modulation coefficient. The calibration basis determines its convergence range to be 1.0 to 1.5. This ratio is a fixed dimensionless value used to characterize the background clutter level of the sensor channel under fault-free operating conditions.

[0078] During actual operation and monitoring, the calibrated preset modulation coefficients will be used... The dynamic occlusion threshold is obtained by multiplying the global pixel variance of the two-dimensional topological feature image generated from the original one-dimensional time series signal. The global pixel variance changes with the image gradient under the current operating conditions, and the dynamic occlusion threshold also changes accordingly. Therefore, the preset modulation coefficient remains at the fixed dimensionless value obtained from the calibration, while the dynamic occlusion threshold is adjusted according to the global gradient distribution of the real-time acquired signal.

[0079] Divide each image into blocks Local two-dimensional gradient variance With dynamic occlusion threshold In the comparison, when the local two-dimensional gradient variance of the image block corresponding to the low-frequency steady-state background outside the diagonal is less than the dynamic occlusion threshold, the image block is downsampled by average pooling and then linearly transformed to generate a single background visual word. When the local two-dimensional gradient variance is not less than the dynamic occlusion threshold, the original feature resolution of the corresponding image block is maintained, and high-frequency abrupt visual words are generated through a linear projection layer. The background visual words and high-frequency abrupt visual words are concatenated according to the spatial topological array order of the original image blocks to form a continuous visual feature word sequence. .

[0080] The equipment ledger management terminal reads the target bearing's equipment operation ledger, cumulative service time, rated life, historical fault mode library, and current operating condition parameters from the historical operation database. It determines the safe service cycle stage based on the ratio of cumulative service time to rated life. When the ratio reaches 0.8, it extracts the frequency domain features corresponding to early inner ring faults and early outer ring faults from the historical fault mode library. It then combines the equipment operation ledger, safe service cycle stage, fault features from the historical fault mode library, and current operating condition parameters into the current operating condition constraint text.

[0081] The word segmenter segments and semantically encodes the current working condition constraint text, generating a text word sequence. Using text word sequences as leading indicators, and combining them with continuous visual feature word sequences... The multimodal context input sequence is formed by concatenating the autoregressive sequences from beginning to end. .

[0082] The system constructs a binarized attention mask matrix based on the index positions of background visual terms in the multimodal context input sequence. In the self-attention computation layer of the large-scale basic inference model with 7B parameters, the attention mask matrix shields the self-attention weight allocation between background visual words, so that the cross-attention computation is focused on the text words and high-frequency mutation visual words. The large-scale basic inference model matches the local texture represented by high-frequency mutation visual words according to the safe service cycle stage, historical failure mode and current operating condition parameters in the current operating condition constraint text, and outputs a structured judgment text containing fault diagnosis labels and causal logical inference chain in an autoregressive manner.

[0083] Example 6: This example calibrates the parameters used before the system goes live, and the accelerometer collects time-domain vibration signal sequences. The sequence contains 1024 discrete sampling points with a fixed sampling frequency of 10000Hz. The signal conversion unit uses continuous wavelet transform to convert the time-domain vibration signal sequence. Mapped to a two-dimensional time-frequency energy image The wavelet basis function is Morlet wavelet, and the scaling factor sequence is used. The value range is from 1 to 64, corresponding to a frequency resolution of 15.625 Hz. This represents the time-domain vibration signal sequence used for calibration. This represents a two-dimensional time-frequency energy image obtained from continuous wavelet transform. This represents the scale factor sequence of the Morlet wavelet.

[0084] Continuous wavelet transform is used for auxiliary calibration before the system goes online. It extracts the energy distribution of the signal in different frequency bands through two-dimensional time-frequency energy images. Online pattern recognition still converts the original one-dimensional time series signal into a two-dimensional topological feature image, and then performs multi-scale spatial feature extraction, image block processing, dynamic occlusion threshold comparison, continuous visual feature word sequence generation, and multimodal context reasoning in sequence. During calibration, the two-dimensional time-frequency energy image and the two-dimensional topological feature image formed by the same signal segment are matched according to their time positions. The phase angle oscillation region in the diagonal neighborhood of the two-dimensional topological feature image is checked by the energy concentration region in the two-dimensional time-frequency energy image, which provides a calibration basis for determining the preset modulation coefficient.

[0085] Under a fault-free baseline condition and three typical fault conditions, 500 two-dimensional time-frequency energy images were acquired respectively. The cumulative distribution function of the normalized gray-level histogram of each image was calculated, and the lower limit quantile of the normalized gray-level values ​​was denoted as . The upper limit quantile is denoted as ,when Take 0.02, When the value is 0.98, the background noise intensity of the image is less than 5%, and the contrast of key time-frequency texture features is improved by 35%. Based on this, the calibration process determines the grayscale cropping range of the two-dimensional time-frequency energy image, so that the time-frequency texture under different working conditions adopts the same grayscale mapping conditions.

[0086] Before deployment, the calibration branch uses an image feature extraction model to process the two-dimensional time-frequency energy image. This model contains a 12-layer attention mechanism structure, with the input feature dimension limited to 512 dimensions, and the output is a recognition and classification probability vector. During the parameter setting phase, the downsampling convolution kernel size is adjusted. and step length Set up comparative test conditions, among which... This indicates the size of the downsampling convolution kernel in the calibration branch. This indicates the corresponding step size. This represents a vector composed of the recognition and classification probabilities of each candidate category. Take 3 and When the value is 2, the feature loss rate is less than 0.8%, and the calculation time is 12.4ms.

[0087] The recognition result verification unit starts from the recognition classification probability vector. Obtain the maximum probability value and the second largest probability value And calculate the confidence index. : ,in, Represents the recognition and classification probability vector The maximum probability value in, This represents the second-highest probability value. The difference between the two is represented by the preset decision threshold, denoted as . ,when Greater than or equal to When, retain the current identification and classification result; when Less than At that time, the process of secondary signal acquisition and image resolution reconstruction is triggered.

[0088] After 1000 consecutive offline operating condition tests and calibrations, when When set to 0.35, the false diagnosis rate of pattern recognition is less than 0.1% and the false negative rate is less than 0.05%. The scale factor sequence, gray quantile, downsampling convolution kernel size, stride and preset decision threshold determined by this result together constitute the calibration parameters before deployment.

[0089] Confidence index The classification results are only used for verification in the pre-deployment calibration branch and are not used as the abnormal confidence level of the online pattern recognition output. During online operation, the system will still concatenate the text word sequence formed by the current working condition constraint text with the continuous visual feature word sequence into a multimodal context input sequence, input it into the large-scale basic inference model, and output a structured judgment text containing pattern classification results, abnormal confidence level and causal logical inference chain. The parameters obtained from the pre-deployment calibration are used to form a baseline reference, so that the system can operate under uniform parameter conditions in the industrial electromagnetic interference and mechanical vibration noise environment.

[0090] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A pattern recognition method based on signal-to-image conversion and large-model inference, characterized in that, Includes the following steps: Step S1: Obtain the original one-dimensional time series signal of the target object and extract the signal segment, calculate the phase angle vector of the signal segment in the polar coordinate system, and map it into a two-dimensional topological feature image using the Gram angle field transformation mechanism; Step S2: Input the two-dimensional topological feature image into the visual encoder to extract multi-scale spatial feature maps, and use a linear projection layer to divide it into non-overlapping image blocks; Step S3: Calculate the local two-dimensional gradient variance and global pixel variance of each image block, determine the dynamic occlusion threshold based on the global pixel variance, perform average pooling downsampling on image blocks with local two-dimensional gradient variance less than the dynamic occlusion threshold and map them as background visual words, and map image blocks with local two-dimensional gradient variance not less than the dynamic occlusion threshold as high-frequency mutation visual words according to the original feature resolution. Step S4: Obtain the current working condition constraint text of the target object and convert it into a text word sequence. Concatenate background visual words and high-frequency mutation visual words according to the sequence dimension to form a continuous visual feature word sequence with non-uniform resolution distribution. Concatenate the text word sequence and the continuous visual feature word sequence to generate a multimodal context input sequence. Input the sequence into a large-scale basic inference model. Calculate the cross-attention weights between high-frequency mutation visual words, background visual words, and text words using a self-attention mechanism. Infer the spatiotemporal pattern evolution law of the two-dimensional topological feature image. Output a structured judgment text containing pattern classification results and causal logical inference chain.

2. The pattern recognition method based on signal-to-image conversion and large-model inference according to claim 1, characterized in that, In step S3, calculating the local two-dimensional gradient variance and global pixel variance of each image block, and determining the dynamic occlusion threshold based on the global pixel variance, includes the following sub-steps: Step S31, calculating the first-order spatial difference of each pixel in the horizontal and vertical directions within each image block to obtain the local two-dimensional spatial pixel gradient of each pixel; Step S32, based on the local two-dimensional spatial pixel gradient of each pixel, calculating the discrete variance of all pixel gradient values ​​within the corresponding image block, as the local two-dimensional gradient variance of the image block; Step S33, summarizing the gradient values ​​of all pixels in the multi-scale spatial feature map, calculating the mean and overall variance of the spatial pixel gradient of the entire image, and determining the dynamic occlusion threshold by multiplying the overall variance by the preset modulation coefficient; Step S34, comparing the local two-dimensional gradient variance of each image block with the dynamic occlusion threshold in real time. When the local two-dimensional gradient variance is lower than the dynamic occlusion threshold, the corresponding image block is determined to belong to the low-frequency steady-state background; when the local two-dimensional gradient variance is not less than the dynamic occlusion threshold, the corresponding image block is determined to belong to the high-frequency transient impact region.

3. The pattern recognition method based on signal-to-image conversion and large-model inference according to claim 2, characterized in that, The process of downsampling and mapping image blocks with local two-dimensional gradient variance less than the dynamic occlusion threshold to background visual words, and mapping image blocks with local two-dimensional gradient variance not less than the dynamic occlusion threshold to high-frequency abrupt visual words according to the original feature resolution, includes the following sub-steps: Step S35: For image blocks determined to belong to low-frequency steady-state background, a two-dimensional spatial average pooling kernel is used to downsample the feature matrix of the image block, compressing the multi-dimensional pixel features into a single low-resolution feature vector, and generating a single background visual word through linear transformation; Step S36: For image blocks determined to belong to high-frequency transient impact regions, the original pixel feature resolution of the image block is maintained, and multiple high-resolution high-frequency abrupt visual words are generated through full mapping by a linear projection layer; Step S37: According to the spatial topological array order of the original image blocks, the background visual words and each high-frequency abrupt visual word are interwoven and spliced ​​to form a continuous visual feature word sequence with a spatially non-uniform resolution distribution.

4. The pattern recognition method based on signal-to-image conversion and large-model inference according to claim 1, characterized in that, Step S4 involves obtaining the current operating condition constraint text of the target object and converting it into a text word sequence, as well as concatenating it with a continuous visual feature word sequence. This includes the following sub-steps: Step S41: Based on a preset structured prompt word template, obtain the target object's equipment operation log, safe service cycle stage, historical fault mode library, and current operating condition parameters, and combine them to construct the current operating condition constraint text; Step S42: Segment and semantically encode the current operating condition constraint text using a preset word segmenter, and map it to generate a text word sequence; Step S43: Use the text word sequence as a pre-guided identifier, and concatenate it with the continuous visual feature word sequence in an autoregressive sequence order, connecting the beginning and end, to generate a multimodal context input sequence containing domain expert prior constraints and image topological details.

5. The pattern recognition method based on signal-to-image conversion and large-model inference according to claim 4, characterized in that, The calculation of cross-attention weights between high-frequency mutation visual words, background visual words, and text words using a self-attention mechanism includes: in the self-attention calculation layer of the large-scale basic inference model, the allocation of self-attention calculation weights between background visual words is shielded, and self-attention calculation resources are concentrated on the calculation of cross-attention correlation between text words and high-frequency mutation visual words; based on the cross-attention correlation calculation results, an isomorphic mapping relationship is established between the semantic vector representing the specific fault evolution trend in the current working condition constraint text and the feature vector representing the violent phase angle oscillation texture in the high-frequency mutation visual words.

6. The pattern recognition method based on signal-to-image conversion and large-model inference according to claim 1, characterized in that, Before obtaining the original one-dimensional time series signal of the target object in step S1, the method further includes: setting the discretization sampling rate to 10kHz according to the Nyquist sampling theorem; normalizing the amplitude of the extracted signal segment and linearly mapping the amplitude to the interval between -1 and 1; setting the time-domain sliding window slice length of the extracted signal segment to 1024 sampling points; and setting the size of each image block obtained by the linear projection layer when segmenting the multi-scale spatial feature map to 16×16 pixels.

7. The pattern recognition method based on signal-to-image conversion and large-model inference according to claim 1, characterized in that, The output includes a structured judgment text containing pattern classification results and a causal logic inference chain. The output includes: a diagnostic classification label representing the current operating state of the target object; and the corresponding causal logic inference chain, which includes a description of the location of high-frequency texture distortion in the two-dimensional topological feature image, an anomaly confidence level calculated by combining the current operating condition constraint text, and a technical causal explanation text for determining the current operating state.

8. The pattern recognition method based on signal-to-image conversion and large-model inference according to claim 6, characterized in that, After normalizing the amplitude of the intercepted signal segments, the process also includes: real-time monitoring of the signal variance volatility of the original one-dimensional time series signal; when the signal variance volatility exceeds a preset volatility threshold, the overlap rate of the time-domain sliding window of the intercepted signal segments is automatically increased from the preset initial overlap rate to the target overlap rate, so as to increase the mapping frequency of the transient impact region in the adjacent time-domain sliding window segments.

9. The pattern recognition method based on signal-to-image conversion and large-model inference according to claim 1, characterized in that, After outputting the structured judgment text containing the pattern classification results and the causal logical deduction chain, the following steps are also included: when the pattern classification results indicate that the target object is in an abnormal fault state, based on the abnormal confidence level and technical causal explanation text output in the structured judgment text, a control protection command is generated; the control protection command is transmitted to the control terminal of the target object, driving the control terminal to adjust the operating power of the target object, and triggering a shutdown response when the abnormal confidence level reaches a preset alarm threshold.

Images