Frequency data analysis method and device based on self-supervised learning, and storage medium

Abstract

The invention discloses a frequency data analysis method and device based on self-supervised learning, and a storage medium. The method comprises the following steps: S1, constructing a standard spectrum tensor; s2, executing frequency change feature extraction operation to generate a frequency change feature map; s3, determining a frequency section of which the hopping intensity is greater than a preset hopping threshold value, constructing a hopping section mask, and generating a structure covering spectrum tensor; s4, inputting to an improved ConvTransform model, and generating a frequency spectrum reconstruction tensor and a residual feature tensor; s5, constructing a frequency spectrum reconstruction loss function and a residual error fitting loss function, executing a self-supervised training process, and constructing a frequency guidance feedback graph; and S6, applying to frequency data analysis, and generating a frequency spectrum reconstruction result, a residual error response result and a hopping section positioning result. According to the invention, the abnormal change identification capability and the frequency spectrum reconstruction precision in the frequency data analysis process are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and frequency domain signal processing technology, and in particular to a frequency data analysis method and device based on self-supervised learning and a storage medium. BACKGROUND

[0002] With the growing demand for vibration anomaly monitoring, power system state evaluation and electroencephalogram pattern recognition, deep analysis and high-precision modeling of frequency domain signals have gradually become a key link in intelligent sensing systems. Existing frequency data analysis methods mainly rely on feature vectors converted from frequency spectrum as model input, and use standard supervised learning models for classification or regression, but still face many challenges in practical applications.

[0003] Due to the interference of factors such as sampling noise, window truncation and edge energy diffusion during the construction of frequency data, the frequency band in the frequency spectrum is prone to be blurred and key information is lost, which limits the feature extraction capability of the model at the frequency level; traditional methods generally rely on manually labeled data for training, which makes it difficult to obtain a large number of labeled frequency data samples in actual scenarios, limiting the generalization ability and adaptability of the algorithm; in addition, the modeling of the frequency spectrum structure generally lacks reconstruction and verification mechanism, making it difficult to realize the semantic consistency judgment between frequency spectra, resulting in poor stability of the frequency domain recognition result and serious feature drift. Especially in the scene where the label is missing or the sample distribution is uneven, the model is prone to overfitting, making it difficult to effectively identify unknown patterns.

[0004] Therefore, how to provide a frequency data analysis method, device and storage medium based on self-supervised learning is a problem that needs to be solved by those skilled in the art. SUMMARY

[0005] One object of the present application is to provide a frequency data analysis method, device and storage medium based on self-supervised learning, which fully utilizes the frequency spectrum tensor construction, frequency change feature map extraction, jump section mask generation and improved ConvTransformer structure, and describes in detail the whole process of generating standard frequency spectrum tensor and frequency change feature map based on frequency data, constructing jump section mask and context position encoding, performing structure masking and frequency spectrum reconstruction, constructing frequency spectrum reconstruction loss function and residual fitting loss function and joint training, finally outputting frequency spectrum reconstruction tensor, residual feature tensor, jump section positioning result and frequency guide feedback graph, which has the advantages of no manual labeling, strong self-supervised training ability, high jump recognition accuracy and strong frequency spectrum anomaly positioning ability.

[0006] The frequency data analysis method based on self-supervised learning according to an embodiment of the present application comprises the following steps:

[0007] S1, collect frequency data and perform denoising processing, normalization processing and time alignment processing to construct a standard spectral tensor;

[0008] S2, perform frequency change feature extraction operation on the standard spectral tensor, generate frequency change feature map based on first-order change quantity, second-order change quantity and local energy change quantity;

[0009] S3, according to the frequency change feature map, determine the frequency section with jump strength greater than the preset jump threshold, construct the jump section mask, and set the remaining frequency section to the mask ratio to perform structure masking processing, generate a structure masked spectral tensor;

[0010] S4, input the structure masked spectral tensor into the improved ConvTransformer model to perform spectral reconstruction processing and residual generation processing, generate spectral reconstruction tensor and residual feature tensor; the improved ConvTransformer model introduces jump perception attention structure, residual alignment structure and position encoding structure;

[0011] S5, based on the spectral reconstruction tensor and the residual feature tensor, construct the spectral reconstruction loss function and the residual fitting loss function, and generate the joint training target by weighted fusion; perform self-supervised training process, output reconstruction error distribution, residual error distribution and attention response value, construct frequency guided feedback graph; update the jump section mask and the context position encoding according to the frequency guided feedback graph;

[0012] S6, apply the improved ConvTransformer model trained to frequency data analysis to generate spectral reconstruction result, residual response result and jump section positioning result.

[0013] Optionally, the S1 is specifically:

[0014] Collect frequency data, the frequency data includes frequency value and corresponding amplitude information at multiple time points, and construct a frequency sequence set;

[0015] Perform denoising processing on the frequency sequence set, and perform smoothing operation on the local value of the frequency sequence by using sliding window filtering method to generate denoised frequency sequence;

[0016] Perform normalization processing on the denoised frequency sequence, perform linear transformation according to the preset frequency range and amplitude range to generate normalized frequency sequence;

[0017] Perform time alignment processing on the normalized frequency sequence, perform interpolation operation or truncation operation according to uniform time step to generate aligned frequency sequence;

[0018] The aligned frequency sequence is constructed into a three-dimensional tensor form in the frequency dimension, the time dimension and the amplitude dimension to generate a standard spectrum tensor; the standard spectrum tensor comprises the frequency dimension, the time dimension and the amplitude dimension.

[0019] Optionally, S2 is specifically:

[0020] A local sliding operation is performed on the standard spectrum tensor in the frequency dimension and the time dimension to calculate the frequency value difference between adjacent time points to generate a first-order change quantity tensor;

[0021] A difference operation is continuously performed on the first-order change quantity tensor to generate a difference value between adjacent first-order change quantities to construct a second-order change quantity tensor;

[0022] Amplitude energy information is extracted from the standard spectrum tensor in a local frequency window and a time window range to calculate the sum of squares of amplitudes of each frequency point in the specified window to generate a local energy change quantity tensor;

[0023] The first-order change quantity tensor, the second-order change quantity tensor and the local energy change quantity tensor are aligned and spliced in the frequency dimension to construct a frequency change feature map.

[0024] Optionally, S3 is specifically:

[0025] A threshold judgment operation is performed on the jump strength value corresponding to each frequency point in the frequency change feature map to filter a frequency section with a jump strength greater than a preset jump threshold, which is marked as a jump frequency section;

[0026] A jump section mask corresponding to the jump frequency section is constructed, and the jump frequency positions in the jump section mask are marked as masked and reserved, and the remaining frequency positions are marked as blank to be masked;

[0027] In the frequency section outside the jump section mask, a masking operation is randomly performed on the frequency positions according to a preset masking ratio to generate a complete structure masking mask;

[0028] The structure masking mask and the standard spectrum tensor are subjected to corresponding masking processing in the frequency dimension and the time dimension to generate a structure masking spectrum tensor; the structure masking spectrum tensor retains the jump frequency section and masks the non-critical frequency region.

[0029] Optionally, S4 is specifically:

[0030] The structure masking spectrum tensor is input into a convolution processing unit in the frequency dimension and the time dimension to perform local frequency feature extraction processing to generate an initial frequency feature tensor;

[0031] The multi-head attention processing is performed on the initial frequency feature tensor, an attention weight matrix is constructed, a jump intensity matrix is introduced in each attention head, the attention distribution is adjusted according to the jump intensity value, and a jump perception feature tensor is generated;

[0032] The feedforward mapping processing and the normalization processing are performed on the jump perception feature tensor, and a context fusion feature tensor is generated.

[0033] A frequency position encoding vector is constructed, a context position encoding tensor is generated according to the frequency index and the context statistical quantity of the corresponding frequency section, the context position encoding tensor is spliced with the context fusion feature tensor, and a fusion representation tensor is generated.

[0034] The decoding processing is performed on the fusion representation tensor, the original tensor structure is restored according to the frequency dimension and the time dimension, and a spectrum reconstruction tensor is generated.

[0035] The point-by-point difference operation is performed on the spectrum reconstruction tensor and the unmasked section in the structure masked spectrum tensor in the frequency dimension, and a residual feature tensor is generated.

[0036] Optionally, the improved ConvTransformer model introduces a jump perception attention structure, a residual alignment structure and a position encoding structure, specifically:

[0037] The jump perception attention structure constructs a jump intensity matrix, takes the jump intensity value corresponding to each frequency point in the frequency change feature map as a weighting factor, performs element-level weighting operation with the standard attention weight matrix in each attention head, generates a jump perception attention weight matrix, and the jump perception attention weight matrix is used to adjust the attention correlation degree between different frequency positions, and generates a jump perception feature tensor.

[0038] The residual alignment structure performs point-by-point difference operation based on the amplitude difference between the spectrum reconstruction tensor and the unmasked position in the structure masked spectrum tensor, generates a residual feature tensor, and the residual feature tensor and the spectrum reconstruction tensor are jointly input as the output result of the training stage into the loss function calculation process.

[0039] The position encoding structure generates a frequency position encoding vector based on the frequency index position and the context statistical feature, constructs the frequency position encoding vector into a context position encoding tensor, and splices the context position encoding tensor with the jump perception feature tensor to form a fusion representation tensor.

[0040] Optionally, the S5 is specifically:

[0041] A spectrum reconstruction loss function is constructed based on the amplitude difference between the spectrum reconstruction tensor and the unmasked frequency position in the structure masked spectrum tensor, and the spectrum reconstruction loss function calculates the error distribution between the predicted spectrum and the original spectrum according to the frequency dimension and the time dimension.

[0042] constructing a residual fitting loss function based on actual error difference values of the unmasked frequency positions in the residual feature tensor and the structure-masked spectrum tensor, the residual fitting loss function calculating a fitting degree between the residual response and the amplitude error point by point according to the frequency positions;

[0043] performing fusion processing on the spectrum reconstruction loss function and the residual fitting loss function according to a set weighting coefficient to generate a joint training target, the joint training target being used to optimize the spectrum reconstruction path and the residual estimation path;

[0044] performing a self-supervised training process based on the joint training target to update the model parameters and output the reconstruction error distribution, the residual error distribution and the attention response value;

[0045] performing joint mapping processing on the reconstruction error distribution and the residual error distribution in the frequency dimension, and constructing a frequency-guided feedback graph in combination with the attention response value;

[0046] updating the jump section mask according to the guidance score of each frequency position in the frequency-guided feedback graph, adjusting the mask marking of the high-response frequency section, and updating the context position encoding vector of the corresponding position.

[0047] Optionally, the S6 specifically comprises:

[0048] receiving the structure-masked spectrum tensor as input by using the improved ConvTransformer model trained, freezing the parameters of the improved ConvTransformer model in the inference process, performing forward calculation operation according to the frequency dimension and the time dimension, and outputting the spectrum reconstruction tensor;

[0049] performing amplitude difference calculation on the spectrum reconstruction tensor and the corresponding unmasked frequency positions in the structure-masked spectrum tensor to generate the residual feature tensor and construct the residual response result;

[0050] extracting the high-residual frequency section according to the residual amplitude distribution of each frequency position in the residual response result, combining the jump-aware attention response graph generated in the training stage, and calculating a joint jump score index;

[0051] selecting the frequency section with a score higher than a jump positioning threshold based on the joint jump score index, marking it as a jump candidate section, and generating a jump section positioning result;

[0052] outputting the spectrum reconstruction tensor as the spectrum reconstruction result, outputting the residual feature tensor as the residual response result, and outputting the jump candidate section as the jump section positioning result.

[0053] Optionally, the frequency data analysis device based on self-supervised learning comprises the following modules:

[0054] a spectrum processing module, configured to collect frequency data and perform denoising processing, normalization processing and time alignment processing, and construct a standard spectrum tensor;

[0055] a feature extraction module, configured to perform frequency change feature extraction operation on the standard spectrum tensor, and generate a frequency change feature map based on a first-order change amount, a second-order change amount and a local energy change amount;

[0056] a jump construction module, configured to determine a frequency section with a jump strength greater than a preset jump threshold according to the frequency change feature map, construct a jump section mask, and set a masking ratio for the remaining frequency sections to perform structural masking processing, and generate a structural masking spectrum tensor;

[0057] a reconstruction residual module, configured to input the structural masking spectrum tensor into an improved ConvTransformer model to perform spectrum reconstruction processing and residual generation processing, and generate a spectrum reconstruction tensor and a residual feature tensor; the improved ConvTransformer model introduces a jump-aware attention structure, a residual alignment structure and a position encoding structure;

[0058] a loss construction module, configured to construct a spectrum reconstruction loss function and a residual fitting loss function based on the spectrum reconstruction tensor and the residual feature tensor, generate a joint training target through weighted fusion, perform a self-supervised training process, output a reconstruction error distribution, a residual error distribution and an attention response value, and construct a frequency-guided feedback graph; and update the jump section mask and the context position encoding according to the frequency-guided feedback graph;

[0059] an analysis application module, configured to apply the improved ConvTransformer model after training to frequency data analysis, and generate a spectrum reconstruction result, a residual response result and a jump section positioning result.

[0060] Optionally, a frequency data analysis storage medium based on self-supervised learning, the storage medium storing executable instructions, the executable instructions being used for being read by a processor of a computer.

[0061] The present application has the following advantages:

[0062] This invention addresses the issues of frequency fluctuation non-stationarity, missing coverage of key transition segments, and ambiguous distribution of spectrum reconstruction errors in frequency data by constructing frequency change feature maps and generating transition segment masks, combined with structural masking processing and contextual position encoding mechanisms. In the spectrum processing stage, frequency data is collected and a standard spectrum tensor is constructed. In the feature extraction stage, a frequency change feature map containing first-order, second-order, and local energy changes is generated. In the transition construction stage, transition segments are located and a structurally masked spectrum tensor is generated. In the residual reconstruction stage, a transition-aware attention structure, a residual alignment structure, and a position encoding structure are introduced to generate a spectrum reconstruction tensor and a residual feature tensor. In the loss construction stage, the spectrum reconstruction loss function and the residual fitting loss function are combined to output a frequency-guided feedback map and update the transition mask. In the analysis and application stage, inference analysis is performed, outputting spectrum reconstruction results, residual response results, and transition segment location results. Ultimately, this achieves accurate identification of transition segments in frequency data and high-fidelity reconstruction of the spectrum structure, improving the sensitivity of frequency disturbance detection and the efficiency of model self-supervised training. Attached Figure Description

[0063] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0064] Figure 1 This is a flowchart of the frequency data analysis method based on self-supervised learning proposed in this invention;

[0065] Figure 2 This is a schematic diagram of the structure of the improved ConvTransformer model proposed in this invention;

[0066] Figure 3 This is a schematic diagram of the frequency data analysis device based on self-supervised learning proposed in this invention. Detailed Implementation

[0067] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0068] refer to Figures 1-2 The frequency data analysis method based on self-supervised learning includes the following steps:

[0069] S1. Collect frequency data and perform denoising, normalization and time alignment processing to construct a standard spectral tensor;

[0070] S2. Perform frequency change feature extraction operation on the standard spectrum tensor, and generate a frequency change feature map based on the first-order change, second-order change and local energy change.

[0071] S3, determine a frequency section with a jump strength greater than a preset jump threshold according to the frequency change feature map, construct a jump section mask, and perform structural masking processing on the remaining frequency sections with a masking ratio to generate a structural masking spectrum tensor;

[0072] S4, input the structural masking spectrum tensor into an improved ConvTransformer model to perform spectrum reconstruction processing and residual generation processing to generate a spectrum reconstruction tensor and a residual feature tensor; the improved ConvTransformer model introduces a jump-aware attention structure, a residual alignment structure, and a position encoding structure;

[0073] S5, construct a spectrum reconstruction loss function and a residual fitting loss function based on the spectrum reconstruction tensor and the residual feature tensor, generate a joint training target by weighted fusion; perform a self-supervised training process to output a reconstruction error distribution, a residual error distribution, and an attention response value, and construct a frequency-guided feedback map; update the jump section mask and the context position encoding according to the frequency-guided feedback map;

[0074] S6, apply the trained improved ConvTransformer model to frequency data analysis to generate a spectrum reconstruction result, a residual response result, and a jump section positioning result.

[0075] The embodiment improves the consistency of frequency data processing and the stability of model training by constructing a standard spectrum tensor as an input basis; by extracting a frequency change feature map, it strengthens the comprehensive perception ability of first-order change, second-order change and local energy change, and realizes accurate description of frequency jump behavior; further, by constructing a jump section mask and performing structural masking processing on non-jump sections, the sensitivity of the model to local jump information is enhanced and the redundant interference of invalid frequency bands is suppressed; further, by introducing an improved ConvTransformer model with a jump-aware attention structure, a residual alignment structure and a position encoding structure, high-fidelity reconstruction of the masked spectrum tensor and effective extraction of key residual information are realized; at the same time, by jointly training the spectrum reconstruction loss function and the residual fitting loss function, the dual modeling capability of the model for spectrum structure and mutation anomaly is improved; based on the frequency-guided feedback map, the jump section mask and the context position encoding are iteratively updated to optimize the model perception path and feature positioning capability; finally, the trained improved ConvTransformer model is applied to frequency data analysis to output the spectrum reconstruction result, the residual response result and the jump section positioning result, improving the accuracy and generalization ability of frequency data anomaly recognition.

[0076] In the embodiment, S1 is specifically:

[0077] Collect frequency data, the frequency data including frequency values and corresponding amplitude information at multiple time points, and construct a frequency sequence set;

[0078] Perform denoising processing on the frequency sequence set, and perform smoothing operation on local values of the frequency sequence in a sliding window filtering manner to generate a denoised frequency sequence;

[0079] Perform normalization processing on the denoised frequency sequence, and perform linear transformation according to a preset frequency range and amplitude range to generate a normalized frequency sequence;

[0080] Perform time alignment processing on the normalized frequency sequence, and perform interpolation operation or truncation operation according to a uniform time step to generate an aligned frequency sequence;

[0081] Construct the aligned frequency sequence into a three-dimensional tensor form according to the frequency dimension, the time dimension and the amplitude dimension to generate a standard spectrum tensor; the standard spectrum tensor includes the frequency dimension, the time dimension and the amplitude dimension.

[0082] In the embodiment, S2 is specifically:

[0083] Perform local sliding operation on the standard spectrum tensor according to the frequency dimension and the time dimension, calculate the frequency value difference between adjacent time points, and generate a first-order change amount tensor; the local sliding operation is a bidirectional sliding along the frequency dimension and the time dimension under a set step and window size, performs difference calculation on the amplitude change of each frequency point between two continuous time frames to form a first-order change amount;

[0084] Continue to perform difference operation on the first-order change amount tensor to generate difference values between adjacent first-order change amounts, and construct a second-order change amount tensor; the difference operation is mainly in the frequency dimension, and calculates the numerical difference between the first-order change amounts of continuous frequency points, which is used to describe the acceleration characteristics of the change trend in the frequency distribution;

[0085] Extract amplitude energy information according to the local frequency window and the time window range of the standard spectrum tensor, calculate the amplitude square sum of each frequency point in the specified window, and generate a local energy change amount tensor; the amplitude square sum is obtained by selecting a two-dimensional neighborhood region centered at the current frequency point in the spectrum tensor, performing square operation and summation on the amplitude elements, and forming a change amount for representing the local energy set;

[0086] The first-order change amount tensor, the second-order change amount tensor, and the local energy change amount tensor are spliced in the frequency dimension to construct a frequency change feature map; the splicing operation includes performing dimension alignment and uniform mapping on the three change tensors in the frequency dimension, fusing change information of different sources into a multi-channel structure, and facilitating subsequent neural network model to extract joint features; the frequency change feature map is used to enhance the representation ability of weak changes in the frequency domain, improve the resolution of key features in the frequency spectrum, and has the technical effects of distinguishing fine-grained frequency patterns and identifying abnormal frequency changes.

[0087] In the embodiment, the S3 is specifically:

[0088] A threshold judgment operation is performed on the jump intensity value corresponding to each frequency point in the frequency change feature map, frequency sections with jump intensity greater than a preset jump threshold are screened, and are marked as jump frequency sections; the jump intensity value is the absolute change value of the amplitude value of the corresponding frequency between adjacent time points, the amplitude jump is calculated through a sliding time window, and intensity normalization processing is performed to obtain a standardized jump intensity distribution;

[0089] A jump section mask is constructed based on the position of the jump frequency section, the jump section mask has the same dimension structure as the frequency change feature map, and the jump frequency position is marked as hidden and reserved in the frequency dimension, and the remaining frequency positions are marked as blank to be hidden, and all time dimensions remain unchanged; the mask mark is defined in numerical form, the hidden and reserved position is assigned a value of 1, and the position to be hidden is assigned a value of 0;

[0090] In the frequency section outside the jump section mask, a hidden position selection operation is performed according to a preset hidden proportion, the hidden proportion is a percentage value of the total number of non-reserved positions in the frequency dimension; random sampling is performed in each time slice, a hidden position is randomly selected from the frequency points that are not marked as reserved, and is assigned a value of 0 to form a complete structure hidden mask; the complete structure hidden mask includes three states of hidden and reserved, random hidden, and to be hidden in the frequency dimension;

[0091] The complete structure hidden mask and the standard frequency spectrum tensor are subjected to corresponding hidden operations in the frequency dimension and the time dimension, the hidden operation is a bitwise multiplication processing, the spectrum amplitude value of the position with a mask value of 0 is set to zero, the spectrum amplitude value of the position with a mask value of 1 is reserved as the original value, and a structure hidden frequency spectrum tensor is generated; the structure hidden frequency spectrum tensor reserves the spectrum structure corresponding to the jump frequency section and performs random hidden operation on the non-key frequency region;

[0092] The structure hidden frequency spectrum tensor is one of the subsequent abnormal evolution identification input data, effectively enhances the attention ability of the model to the key region of frequency jump, improves the fault tolerance robustness to the low confidence frequency section, realizes the dual technical effects of jump reservation and redundancy suppression.

[0093] In the embodiment, S4 is specifically:

[0094] The structure-masked spectrum tensor is input to a convolution processing unit in the frequency dimension and the time dimension to perform local frequency feature extraction processing; the convolution processing unit includes a plurality of convolution kernel sets with different receptive field sizes for extracting frequency features at different frequency scales; each convolution kernel performs a shared sliding window operation in the time dimension to extract local patterns of frequency changes over time and output an initial frequency feature tensor;

[0095] Multi-head attention processing is performed on the initial frequency feature tensor to generate an attention weight matrix; a jump strength matrix is introduced in each attention head, and the jump strength matrix is calculated based on the amplitude change rate of the edge of the masked section in the structure-masked spectrum tensor, and the amplitude change rate is normalized by the absolute value difference of the amplitude values of adjacent frames in the frequency dimension; the jump strength value is used as an adjustment factor to scale and shift the attention distribution, suppress feature redundancy in smooth regions, enhance jump region perception, and generate a jump perception feature tensor;

[0096] Feedforward mapping processing and normalization processing are performed on the jump perception feature tensor; the feedforward mapping processing includes two layers of fully connected mapping structure with an activation function embedded in the middle to improve the nonlinear expression capability; the normalization processing adopts layer normalization to perform mean and standard deviation normalization operations on the feature values of each frequency position to generate a context fusion feature tensor;

[0097] A frequency position encoding vector is constructed, which is generated based on the frequency index and the context statistics of the corresponding frequency section; the context statistics include the average amplitude, the maximum amplitude, and the jump frequency density, which is obtained by counting the position frequency of the jump strength value exceeding a preset threshold within a unit time; the frequency position encoding vector is extended to the frequency dimension to generate a context position encoding tensor; the context position encoding tensor and the context fusion feature tensor are spliced and combined in the channel dimension to generate a fusion representation tensor;

[0098] Decoding processing is performed on the fusion representation tensor; the decoding processing includes double-channel deconvolution operations in the frequency dimension and the time dimension to restore the structure shape of the original tensor and output a spectrum reconstruction tensor; a jump guide gating structure is introduced in the decoding process to adjust the activation weight of the deconvolution channel according to the jump strength matrix to improve the restoration accuracy of the jump spectrum structure;

[0099] Point-by-point difference operations are performed on the spectrum reconstruction tensor and the unmasked section in the structure-masked spectrum tensor in the frequency dimension to generate a residual feature tensor; the point-by-point difference operation is a direct difference between the reconstructed value and the true value at the corresponding frequency point, and the residual feature tensor is used for subsequent residual fitting processing.

[0100] The embodiment enhances the local perception of the frequency hopping structure in the frequency dimension, improves the recovery accuracy of the mutation area in the reconstruction process, and preserves the context frequency statistical characteristics, thereby improving the expression integrity and accuracy of the spectrum reconstruction tensor.

[0101] In the embodiment, the improved ConvTransformer model introduces a hopping perception attention structure, a residual alignment structure, and a position encoding structure, specifically:

[0102] The hopping perception attention structure takes the hopping intensity value corresponding to each frequency point in the frequency variation feature map as a weighting factor by constructing a hopping intensity matrix, and the hopping intensity value is calculated based on the amplitude difference between consecutive time segments. The hopping intensity value and the attention distribution between each frequency point in the standard attention weight matrix are subjected to point-by-point weighting operation to generate a hopping perception attention weight matrix in each attention head. The hopping perception attention weight matrix is used to adjust the attention correlation degree between different frequency positions, so that the attention mechanism enhances the weight expression at the frequency hopping position, reduces the interference distribution in the frequency stable area, and improves the response ability of the masking recovery process to high dynamic frequency points. The hopping perception attention weight matrix and the frequency variation feature map perform tensor multiplication operation to generate a hopping perception feature tensor.

[0103] The residual alignment structure performs difference operation on each frequency point based on the amplitude difference between the spectrum reconstruction tensor and the unmasked position in the structural masking spectrum tensor to construct a residual value distribution map, which records the amplitude error between the model output and the actual unmasked spectrum. A residual feature tensor is generated according to the residual value distribution map, which is concatenated with the spectrum reconstruction tensor and input to the subsequent loss function calculation path. The joint loss expression is constructed by performing weighted summation operation of the spectrum reconstruction loss function and the residual fitting loss function on the joint residual feature tensor and the spectrum reconstruction tensor, and the joint loss expression is used to reflect the training effect of the model in the two dimensions of spectrum reconstruction accuracy and residual fitting ability.

[0104] The position encoding structure generates a frequency position encoding vector based on the frequency index position and the context statistical characteristics, and the frequency index position is constructed by using discrete position labels, and the context statistical characteristics include local amplitude mean, hopping density and frequency window variance. The frequency position encoding vector is mapped to a context position encoding tensor, which is concatenated with the hopping perception feature tensor in the channel dimension to generate a fusion representation tensor, which is used as the input feature of the subsequent spectrum masking recovery.

[0105] The embodiment effectively improves the expression capability of the model in processing frequency hopping, high residual position and position sensitive area by introducing the jump-aware attention structure, the residual alignment structure and the position encoding structure, realizes high-precision restoration and difference alignment of the structure-occluded spectrum, and enhances the accuracy and robustness of the spectrum reconstruction process.

[0106] In the embodiment, the S5 is specifically:

[0107] A spectrum reconstruction loss function is constructed based on the amplitude difference of the un-occluded frequency positions in the spectrum reconstruction tensor and the structure-occluded spectrum tensor. The spectrum reconstruction loss function performs position-by-position difference operation on the amplitude error between the predicted spectrum tensor and the original spectrum tensor according to the frequency dimension and the time dimension, and constructs a loss expression in the form of L2 norm. In the difference operation, for each position that is not occluded by the structure, the difference between the predicted spectrum amplitude and the original spectrum amplitude is calculated, and all the differences are squared and summed to generate the total spectrum reconstruction error;

[0108] A residual fitting loss function is constructed based on the actual error difference of the un-occluded frequency positions in the residual feature tensor and the structure-occluded spectrum tensor. The residual fitting loss function evaluates the response difference between the residual feature tensor and the spectrum reconstruction error according to the frequency dimension. The residual value at each frequency position is subjected to point-by-point difference operation with the prediction error at the position, and the weighted L1 norm is used to represent the response fitting degree, and the total residual fitting error is output.

[0109] The spectrum reconstruction loss function and the residual fitting loss function are fused according to the set weighting coefficients; two non-zero weighting coefficients are set to linearly weight and sum the total spectrum reconstruction error and the total residual fitting error to generate a joint training target, which represents the comprehensive optimization direction of the spectrum reconstruction path and the residual estimation path.

[0110] A self-supervised training process is performed based on the joint training target; the joint training target is used as the minimization objective function of the model, and the gradient descent method is used for back propagation update to optimize the parameters of each layer of the spectrum prediction module in the spectrum reconstruction path and the parameters of each layer of the error extraction module in the residual estimation path, and finally output the reconstruction error distribution tensor, the residual error distribution tensor and the attention response value vector.

[0111] The reconstruction error distribution tensor and the residual error distribution tensor are jointly mapped in the frequency dimension; for each frequency position, the corresponding reconstruction error value and residual error value are concatenated and input into the frequency fusion mapping structure to output the fusion response vector of each frequency position; the fusion response vector is subjected to attention weighting in combination with the attention response value vector to generate a frequency-guided feedback graph, in which each frequency position records a guide score for evaluating the importance and response strength of the frequency position.

[0112] According to the guide score of each frequency position in the frequency guide feedback graph, a jump section mask is updated; for the position with a guide score greater than a set masking threshold, a masking label enhancement operation is performed, the corresponding position value in the jump section mask is updated to 1, indicating a high response frequency position; and the context position encoding vector of the corresponding position is updated; the part containing the high response frequency in the context position encoding vector is introduced into the high position signal weight, and the original encoding value is adjusted through the frequency position index function, so that the information of the high importance position is explicitly improved.

[0113] The embodiment can enhance the cooperation ability of the spectrum prediction path and the error estimation path, and improve the spectrum reconstruction accuracy and error fitting precision; the construction of the frequency guide feedback graph and the dynamic update of the jump section mask can guide the model to focus on the key frequency position, improve the spectrum structure recovery ability and frequency sensitivity, and improve the precision and robustness of the frequency data anomaly detection.

[0114] In the embodiment, the S6 is specifically:

[0115] The improved ConvTransformer model trained is used to receive a structure masked spectrum tensor as input, and the parameters of the improved ConvTransformer model are frozen in the inference process; a forward calculation operation is performed according to the frequency dimension and the time dimension, the context features of the missing part in the spectrum tensor are extracted through the multi-head attention mechanism and the convolution structure, the local and global spectrum patterns are fused, and a spectrum reconstruction tensor is output;

[0116] The amplitude difference between the spectrum reconstruction tensor and the corresponding unmasked frequency position in the structure masked spectrum tensor is calculated, and the calculation method is to calculate the absolute difference between the reconstruction spectrum value and the original spectrum value in the time sequence for each unmasked frequency point, generate a residual feature tensor, and construct a residual response result;

[0117] According to the residual amplitude distribution of each frequency position in the residual response result, a one-dimensional residual curve of the frequency position and the residual amplitude is constructed; a stable residual section is extracted by performing a sliding average processing on the residual curve, a residual slope is calculated by using a difference operation, a high residual frequency section is identified as a jump sensitive candidate section, and a jump sensitive candidate section is extracted;

[0118] The jump perception attention response graph generated in the training stage is subjected to frequency dimension alignment processing with the extracted jump sensitive candidate section, the attention value of each candidate frequency position in the attention graph is calculated, a joint jump score index is constructed by combining the residual amplitude normalized value, the joint jump score index is a weighted product of the residual amplitude normalized value and the attention value, and the weight proportion is set according to the jump sample distribution experience value obtained by statistical analysis of the training set;

[0119] Screening frequency segments with scores higher than a jump positioning threshold based on the joint jump score index, the jump positioning threshold being a preset percentile threshold, preferably a 95th percentile value of the joint score; taking the screened high-score frequency segments as jump candidate segments to generate a jump segment positioning result;

[0120] Outputting the spectrum reconstruction tensor as a spectrum reconstruction result, the residual feature tensor as a residual response result, and the jump candidate segments as a jump segment positioning result.

[0121] The embodiment can accurately restore the spectral integrity in the case of spectral masking, and accurately position the jump segments by combining residual changes and attention responses, thereby achieving high spectrum reconstruction accuracy and strong jump detection sensitivity.

[0122] Reference Figure 3 In the embodiment, the frequency data analysis device based on self-supervised learning includes the following modules:

[0123] The spectrum processing module is configured to collect frequency data and perform denoising, normalization and time alignment processing to construct a standard spectrum tensor;

[0124] The feature extraction module is configured to perform frequency change feature extraction on the standard spectrum tensor to generate a frequency change feature map based on first-order change, second-order change and local energy change;

[0125] The jump construction module is configured to determine frequency segments with jump intensity greater than a preset jump threshold based on the frequency change feature map, construct a jump segment mask, and set a masking ratio for the remaining frequency segments to perform structural masking processing to generate a structural masking spectrum tensor;

[0126] The reconstruction residual module is configured to input the structural masking spectrum tensor into an improved ConvTransformer model to perform spectrum reconstruction processing and residual generation processing to generate a spectrum reconstruction tensor and a residual feature tensor; the improved ConvTransformer model introduces a jump-aware attention structure, a residual alignment structure and a position encoding structure;

[0127] The loss construction module is configured to construct a spectrum reconstruction loss function and a residual fitting loss function based on the spectrum reconstruction tensor and the residual feature tensor, generate a joint training target by weighted fusion, perform a self-supervised training process to output a reconstruction error distribution, a residual error distribution and an attention response value, and construct a frequency-guided feedback graph; and update the jump segment mask and the context position encoding according to the frequency-guided feedback graph;

[0128] The analysis application module is configured to apply the improved ConvTransformer model to frequency data analysis to generate a frequency spectrum reconstruction result, a residual response result, and a jump section positioning result.

[0129] In the embodiment, a frequency data analysis storage medium based on self-supervised learning stores executable instructions for being read by a processor of a computer.

[0130] Embodiment 1

[0131] To verify the feasibility of the application in implementation, the application is applied to a high-frequency sampling electrical equipment state monitoring system of a power substation. The substation deploys several high-frequency current, voltage, vibration, and sound signal acquisition devices, and each device generates more than 50 GB of frequency data per day. Most of the data lack explicit labels and only have limited fault annotations and alarm logs. In the traditional analysis method, only a small amount of labeled data can be used for model training, resulting in delayed fault discovery, low frequency feature extraction accuracy, and difficulty in meeting real-time and comprehensive monitoring needs.

[0132] In this scenario, the frequency data analysis method based on self-supervised learning is combined to obtain the original frequency signal through the frequency data acquisition module. The frequency spectrum and the spectrum graph are generated by performing frequency spectrum transformation, and the frequency subtask set is automatically generated through the subtask construction module. The frequency prediction subtask, the mask reconstruction subtask, and the change detection subtask are constructed. Further, the frequency spectrum contrast features and the spectrum structure features are extracted based on the feature contrast learning module, the frequency time series feature vector is constructed, and input into the fusion discrimination module. The signal state classification and change trend discrimination are performed, and the frequency data analysis result containing the confidence score is finally output.

[0133] In actual deployment, three power equipment, a circuit breaker (device A), a main transformer (device B), and a reactor (device C), are selected as the main monitoring targets, and the application method is compared with the traditional supervised convolution model. The continuous monitoring time is three months, and the key indicators such as frequency anomaly recognition accuracy, average detection time, unlabeled data utilization rate, and false alarm rate are evaluated. The following is a comparison data table:

[0134] Table 1 Performance comparison table of different methods in frequency data anomaly detection

[0135]

[0136] From the data in Table 1, it can be seen that in the frequency data anomaly detection process of the three devices, the method of the present application is significantly better than the traditional supervised method in recognition accuracy. The anomaly recognition accuracy is improved by about 11.8%, which reflects the strong generalization ability of the frequency spectrum contrast feature and the spectral series structure feature in the unlabelled environment. In terms of average detection time, the traditional method relies on static convolutional network and responds slowly to high-frequency data, and it takes an average of 412 milliseconds to complete one detection; in the execution of the subtask process, the present application uses the mask reconstruction and change detection mechanism to quickly locate the frequency fluctuation segment, shortening the detection time to 185 milliseconds, and has stronger real-time response capability. In addition, the utilization rate of unlabelled data is one of the core advantages of the present application, and the traditional method can only utilize less than 30% of the data, while the method of the present application can utilize more than 90% of the data through subtask design and self-supervised training mechanism. In terms of false positive rate control, the traditional model often misjudges due to frequency similarity interference, while the present application realizes more stable decision-making by fusing confidence scores, and the false positive rate is significantly reduced to 3.1%.

[0137] In order to further analyze the stability of the method at the device level, the monthly average fault recognition and false positive number of devices A, B and C during the three months are selected for comparison, and the results are shown in Table 2.

[0138] Table 2 Comparison of fault recognition statistics of three types of devices

[0139]

[0140] From the data in Table 2, it can be observed that the present application has higher fault recognition ability and lower false positive and missed detection in the three devices. Taking device A as an example, the traditional method only correctly identifies 14 faults in three months, with 5 false positives and 3 missed detections; while the method of the present application can achieve 18 accurate recognitions, with only 1 false positive and 0 missed detection rate, which reflects the sufficient extraction ability of frequency subtask learning to real fault features. Similarly, in devices B and C, the present application also shows the same trend, achieving zero missed detection and low false positive. This stability is extremely critical for the safety monitoring of actual running devices, and can effectively improve the credibility of the early warning system.

[0141] In the high-frequency signal scene lacking labels, the present embodiment can greatly improve the accuracy, response speed and generalization ability of frequency anomaly recognition, and has strong unlabelled adaptive modeling ability and decision robustness. Compared with the traditional model relying on manual annotation and fixed mode, the present application effectively reduces the dependence on training data and false positive cost, improves the level of automatic frequency spectrum monitoring of actual industrial systems, and is suitable for frequency data intelligent analysis scenes in multiple key fields such as power, manufacturing, transportation and energy.

[0142] The above merely describes preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art, according to the technical solution and inventive concept of the present application, makes equivalent replacement or change within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for frequency data analysis based on self-supervised learning, characterized in that, The method comprises the following steps: S1, collecting frequency data and performing denoising processing, normalization processing and time alignment processing to construct a standard spectral tensor; S2, performing frequency change feature extraction operation on the standard spectral tensor, generating frequency change feature map based on first-order change quantity, second-order change quantity and local energy change quantity; S3, according to the frequency change feature map, determining the frequency section with jump strength greater than the preset jump threshold, constructing the jump section mask, and setting the masking ratio for the remaining frequency section to perform structure masking processing, generating a structure masked spectral tensor; S4, inputting the structure masked spectral tensor into the improved ConvTransformer model to perform spectral reconstruction processing and residual generation processing, generating a spectral reconstruction tensor and a residual feature tensor; the improved ConvTransformer model introduces a jump perception attention structure, a residual alignment structure and a position encoding structure; S5, based on the spectral reconstruction tensor and the residual feature tensor, constructing a spectral reconstruction loss function and a residual fitting loss function, and generating a joint training target by weighted fusion; performing a self-supervised training process to output reconstruction error distribution, residual error distribution and attention response value, and constructing a frequency guided feedback graph; updating the jump section mask and the context position encoding according to the frequency guided feedback graph; S6, applying the trained improved ConvTransformer model to frequency data analysis to generate spectral reconstruction results, residual response results and jump section positioning results.

2. The self-supervised learning based frequency data analysis method of claim 1, wherein, The S1 is specifically: Collecting frequency data, the frequency data including frequency values and corresponding amplitude information at multiple time points to construct a frequency sequence set; Performing denoising processing on the frequency sequence set, using a sliding window filtering method to smooth the local value of the frequency sequence to generate a denoised frequency sequence; Performing normalization processing on the denoised frequency sequence, performing linear transformation according to the preset frequency range and amplitude range to generate a normalized frequency sequence; Performing time alignment processing on the normalized frequency sequence, performing interpolation operation or truncation operation according to a uniform time step to generate an aligned frequency sequence; Constructing the aligned frequency sequence into a three-dimensional tensor form according to the frequency dimension, time dimension and amplitude dimension to generate a standard spectral tensor; the standard spectral tensor includes frequency dimension, time dimension and amplitude dimension. 3.The self-supervised learning based frequency data analysis method of claim 1, wherein, The S2 is specifically: Performing local sliding operation on the standard spectral tensor according to the frequency dimension and the time dimension, calculating the frequency value difference between adjacent time points to generate a first-order change quantity tensor; Continuing to perform difference operation on the first-order change quantity tensor to generate difference values between adjacent first-order change quantities, constructing a second-order change quantity tensor; Extracting amplitude energy information according to the local frequency window and the time window range of the standard spectral tensor, calculating the sum of amplitude squares of each frequency point in the specified window to generate a local energy change quantity tensor; Aligning and splicing the first-order change quantity tensor, the second-order change quantity tensor and the local energy change quantity tensor according to the frequency dimension to construct a frequency change feature map. 4.The self-supervised learning based frequency data analysis method of claim 1, wherein, The S3 is specifically: Perform a threshold judgment operation on the jump strength value corresponding to each frequency point in the frequency change feature map, filter the frequency section with jump strength greater than the preset jump threshold, and mark it as a jump frequency section; A jump section mask corresponding to the jump frequency section is constructed, and the jump frequency position in the jump section mask is marked as a mask reservation, and the remaining frequency positions are marked as blank to be masked; In the frequency section outside the jump section mask, a frequency position is randomly selected according to a preset masking ratio to perform a masking operation, and a complete structure mask is generated; The structure mask and the standard spectral tensor are correspondingly masked in the frequency dimension and the time dimension to generate a structure mask spectral tensor; the structure mask spectral tensor retains the jump frequency section and masks the non-critical frequency area.

5. The self-supervised learning based frequency data analysis method of claim 1, wherein, The S4 specifically is: The structure mask spectral tensor is input into the convolution processing unit according to the frequency dimension and the time dimension, local frequency feature extraction processing is performed, and an initial frequency feature tensor is generated; A multi-head attention processing is performed on the initial frequency feature tensor, an attention weight matrix is constructed, a jump strength matrix is introduced in each attention head, the attention distribution is adjusted according to the jump strength value, and a jump perception feature tensor is generated; A feedforward mapping processing and a normalization processing are performed on the jump perception feature tensor to generate a context fusion feature tensor; A frequency position encoding vector is constructed, a context position encoding tensor is generated according to the frequency index and the context statistical quantity of the corresponding frequency section, the context position encoding tensor and the context fusion feature tensor are spliced to generate a fusion representation tensor; A decoding processing is performed on the fusion representation tensor, the original tensor structure is recovered according to the frequency dimension and the time dimension, and a spectral reconstruction tensor is generated; The spectral reconstruction tensor and the unmasked section in the structure mask spectral tensor are point-by-point difference operated in the frequency dimension to generate a residual feature tensor.

6. The self-supervised learning based frequency data analysis method of claim 1, wherein, The improved ConvTransformer model introduces a jump perception attention structure, a residual alignment structure and a position encoding structure, specifically: The jump perception attention structure constructs a jump strength matrix, takes the jump strength value corresponding to each frequency point in the frequency change feature map as a weighting factor, and performs element-level weighting operation with the standard attention weight matrix in each attention head to generate a jump perception attention weight matrix, which is used to adjust the attention correlation degree between different frequency positions and generate a jump perception feature tensor; The residual alignment structure performs point-by-point difference operation based on the amplitude difference between the spectral reconstruction tensor and the unmasked position in the structure mask spectral tensor to generate a residual feature tensor, which is input into the loss function calculation process together with the spectral reconstruction tensor as the output result in the training stage; The position encoding structure generates a frequency position encoding vector based on the frequency index position and the context statistical feature, constructs a context position encoding tensor from the frequency position encoding vector, and splices the context position encoding tensor with the jump perception feature tensor to form a fusion representation tensor.

7. The self-supervised learning based frequency data analysis method of claim 1, wherein, The S5 specifically is: constructing a spectrum reconstruction loss function based on the amplitude difference between the spectrum reconstruction tensor and the unmasked frequency position in the structure-masked spectrum tensor, the spectrum reconstruction loss function calculating the error distribution between the predicted spectrum and the original spectrum according to the frequency dimension and the time dimension; constructing a residual fitting loss function based on the actual error difference between the residual feature tensor and the unmasked frequency position in the structure-masked spectrum tensor, the residual fitting loss function calculating the fitting degree between the residual response and the amplitude error point by point according to the frequency position; performing fusion processing on the spectrum reconstruction loss function and the residual fitting loss function according to the set weighting coefficient to generate a joint training target, the joint training target being used to optimize the spectrum reconstruction path and the residual estimation path; performing a self-supervised training process based on the joint training target to update the model parameters and output the reconstruction error distribution, the residual error distribution and the attention response value; performing joint mapping processing on the reconstruction error distribution and the residual error distribution in the frequency dimension, and constructing a frequency-guided feedback graph combining the attention response value; updating the jump section mask according to the guide score of each frequency position in the frequency-guided feedback graph, adjusting the mask marking of the high-response frequency section, and updating the context position encoding vector of the corresponding position. 8.The self-supervised learning based frequency data analysis method of claim 1, wherein, The S6 specifically comprises: receiving the structure-masked spectrum tensor as input by using the improved ConvTransformer model after training, freezing the parameters of the improved ConvTransformer model in the inference process, performing forward calculation operation according to the frequency dimension and the time dimension, and outputting the spectrum reconstruction tensor; performing amplitude difference calculation on the spectrum reconstruction tensor and the corresponding unmasked frequency position in the structure-masked spectrum tensor to generate the residual feature tensor and construct the residual response result; extracting the high-residual frequency section according to the residual amplitude distribution of each frequency position in the residual response result, combining the jump-aware attention response graph generated in the training stage, and calculating the joint jump score index; selecting the frequency section with a score higher than the jump positioning threshold based on the joint jump score index, marking it as a jump candidate section, and generating a jump section positioning result; outputting the spectrum reconstruction tensor as the spectrum reconstruction result, outputting the residual feature tensor as the residual response result, and outputting the jump candidate section as the jump section positioning result.

9. The apparatus for analyzing frequency data based on self-supervised learning according to any one of claims 1 to 8, wherein the method for analyzing frequency data based on self-supervised learning is executed. The following modules are included: a spectrum processing module for collecting frequency data and performing denoising processing, normalization processing and time alignment processing to construct a standard spectrum tensor; a feature extraction module for performing frequency change feature extraction operation on the standard spectrum tensor, generating a frequency change feature map based on the first-order change amount, the second-order change amount and the local energy change amount; a jump construction module for determining the frequency section with a jump intensity greater than a preset jump threshold according to the frequency change feature map, constructing a jump section mask, and performing structure masking processing on the remaining frequency sections with a masking ratio to generate a structure-masked spectrum tensor; The reconstruction residual module is configured to input the structure-occluded spectral tensor into an improved ConvTransformer model, perform spectral reconstruction processing and residual generation processing, and generate a spectral reconstruction tensor and a residual feature tensor; the improved ConvTransformer model introduces a jump perception attention structure, a residual alignment structure, and a position encoding structure; The loss construction module is configured to construct a spectral reconstruction loss function and a residual fitting loss function based on the spectral reconstruction tensor and the residual feature tensor, generate a joint training target through weighted fusion, perform a self-supervised training process, output a reconstruction error distribution, a residual error distribution, and an attention response value, and construct a frequency-guided feedback graph; and update a jump section mask and a context position encoding according to the frequency-guided feedback graph. The analysis application module is configured to apply the trained improved ConvTransformer model to frequency data analysis, and generate a spectral reconstruction result, a residual response result, and a jump section positioning result.

10. A storage medium for frequency data analysis based on self-supervised learning, characterized in that, The storage medium stores executable instructions, and the executable instructions are used to be read by a processor of a computer to execute the method of claim 1-8.

Images