Method for detecting insulating property of cold plate of power battery and related equipment
By acquiring ultrasonic echo signals from the surface of the cold plate of the power battery, generating multi-dimensional feature maps and performing feature fusion, the problem of low detection accuracy of the insulation layer of the cold plate of the power battery is solved, achieving efficient and accurate insulation performance evaluation and ensuring the safety of the battery system.
Patent Information
- Application Number
- CN202510937189.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-08
- Publication Date
- 2025-11-28
AI Technical Summary
In existing technologies, the detection accuracy and efficiency of the insulation layer of the cold plate in power batteries are low, and it is impossible to fully cover the surface of the cold plate, resulting in inaccurate insulation performance evaluation and difficulty in meeting the high requirements of power battery systems.
By acquiring ultrasonic echo signals from target detection points on the surface of the power battery cold plate, frequency domain, time domain, and time-frequency energy feature maps are generated. By combining the map feature extraction network with the insulation layer classification network, multi-dimensional feature analysis and fusion are achieved, thereby improving detection accuracy and efficiency.
It enables precise capture of the insulation layer thickness and uniformity, improving detection accuracy and efficiency, and ensuring the safe and stable operation of the battery system.
Smart Images

Figure CN121027335A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of power batteries, in particular to a detection method for the insulation performance of a power battery cold plate and related equipment. BACKGROUND
[0002] As a core component of the battery thermal management system, the surface of the power battery cold plate is sprayed with an insulating layer, which is crucial to ensuring the electrical insulation performance, mechanical strength and service life of the battery system. The thickness and uniformity of the insulating layer directly affect the insulation resistance, breakdown risk and coating mechanical strength. If the detection is not accurate, it may cause current leakage, circuit short circuit and other safety hazards.
[0003] In the prior art, manufacturers mainly use a film thickness gauge to manually select points on the surface of the cold plate to detect the thickness of the insulating layer. This method relies on manual subjective operation and has the problems of low detection accuracy and low efficiency, and it is difficult to fully cover the surface of the cold plate and accurately evaluate the overall uniformity of the insulating layer, thereby resulting in an unguaranteed product yield and difficulty in meeting the high requirements of the power battery system on the insulation performance. Therefore, there is an urgent need for a detection method for the insulation performance of a power battery cold plate to solve the above-mentioned technical problems. SUMMARY
[0004] A series of simplified concepts are introduced in the summary section, which will be further described in detail in the specific embodiments section. The summary section of the present application does not mean to attempt to limit the key features and necessary technical features of the claimed technical solution, nor to attempt to determine the protection scope of the claimed technical solution.
[0005] The present application aims to solve the problems of low detection accuracy and low efficiency in the prior art. By integrating different dimensional signal features through multi-modal feature fusion, the feature representation capability is enhanced, the differences in the thickness and uniformity of the insulating layer are captured, and the accuracy and efficiency of the insulating layer detection are improved.
[0006] In a first aspect, the present application provides a detection method for the insulation performance of a power battery cold plate, comprising:
[0007] acquiring an ultrasonic echo signal of a target detection point on the surface of the power battery cold plate;
[0008] generating a frequency domain feature spectrum, a time domain feature spectrum and a time-frequency energy feature spectrum of the target detection point based on the ultrasonic echo signal of the target detection point;
[0009] extracting deep features corresponding to the frequency domain feature spectrum, the time domain feature spectrum and the time-frequency energy feature spectrum of the target detection point based on a spectrum feature extraction network, to obtain a frequency domain feature vector, a time domain feature vector and an energy feature vector of the target detection point;
[0010] The frequency domain feature vector, the time domain feature vector and the energy feature vector of the target detection point are fused to generate a fusion feature vector of the target detection point;
[0011] The fusion feature vector of the target detection point is input into the insulating layer classification network to output an insulating layer detection result of the target detection point;
[0012] Based on the insulating layer detection result of the target detection point, an insulating performance detection result of the power battery cold plate is determined.
[0013] In some embodiments, based on the ultrasonic echo signal of the target detection point, a frequency domain feature spectrum of the target detection point is generated, including:
[0014] The ultrasonic echo signal of the target detection point is subjected to frequency domain transformation to generate an initial frequency spectrum of the target detection point;
[0015] Based on the initial frequency spectrum of the target detection point and a target frequency band, the target frequency band is subjected to frequency spectrum refinement to generate an enhanced frequency spectrum of the target detection point;
[0016] Based on the enhanced frequency spectrum of the target detection point, a frequency domain feature matrix of the target detection point is constructed;
[0017] The frequency domain feature matrix of the target detection point is subjected to graphical conversion to generate a frequency domain feature spectrum of the target detection point.
[0018] In some embodiments, based on the ultrasonic echo signal of the target detection point, a time domain feature spectrum of the target detection point is generated, including:
[0019] Based on the ultrasonic echo signal of the target detection point, a polar coordinate mapping point set of the target detection point is constructed;
[0020] Based on the polar coordinate mapping point set of the target detection point, a symmetric point distribution diagram of the target detection point is generated, and the symmetric point distribution diagram is taken as a time domain feature spectrum of the target detection point.
[0021] In some embodiments, based on the ultrasonic echo signal of the target detection point, a time-frequency energy feature spectrum of the target detection point is generated, including:
[0022] The ultrasonic echo signal of the target detection point is subjected to time-frequency transformation to generate a time-frequency spectrum energy matrix of the target detection point;
[0023] The time-frequency spectrum energy matrix of the target detection point is subjected to frequency band decomposition through a preset filter bank to extract an energy distribution feature of the target detection point;
[0024] Based on the energy distribution feature of the target detection point, a time-frequency energy distribution matrix of the target detection point is constructed;
[0025] convert the time-frequency energy distribution matrix of the target detection point into a visualized heat map, and use the visualized heat map as the time-frequency energy feature map of the target detection point.
[0026] In some embodiments, based on the ultrasonic echo signal of the target detection point, a polar coordinate mapping point set of the target detection point is constructed, including:
[0027] Based on the ultrasonic echo signal of the target detection point, the signal maximum value, the signal minimum value and the current time point amplitude of the target detection point are determined;
[0028] Based on the preset lag factor, the lag time point amplitude of the target detection point is determined;
[0029] Based on the signal maximum value and the signal minimum value, the current time point amplitude of the target detection point is normalized to generate a normalized polar radius of the target detection point;
[0030] Based on the signal maximum value and the signal minimum value, the lag time point amplitude of the target detection point is normalized to generate a lag time point amplitude normalization of the target detection point;
[0031] Based on the lag time point amplitude normalization, the preset reference angle and the preset angle gain factor, the lag time point amplitude of the target detection point is symmetrically deflected to generate a first deflection angle and a second deflection angle of the target detection point;
[0032] Based on the normalized polar radius, the first deflection angle and the second deflection angle, a polar coordinate mapping point set of the target detection point is constructed.
[0033] In some embodiments, the frequency domain feature vector, the time domain feature vector and the energy feature vector of the target detection point are subjected to feature fusion processing to generate a fusion feature vector of the target detection point, including:
[0034] The frequency domain feature vector, the time domain feature vector and the energy feature vector of the target detection point are subjected to feature splicing along the channel dimension to generate a spliced feature vector of the target detection point;
[0035] The spliced feature vector is subjected to convolution to generate a convolution feature vector of the target detection point;
[0036] The convolution feature vector is subjected to feature weighting based on a self-attention mechanism to generate a weighted feature vector of the target detection point;
[0037] The weighted feature vector and the convolution feature vector are subjected to residual fusion to generate a fusion feature vector of the target detection point.
[0038] In some embodiments, the frequency domain feature vector, the time domain feature vector and the energy feature vector of the target detection point are subjected to feature fusion processing to generate a fusion feature vector of the target detection point, including:
[0039] calculate the cosine similarity between each two of the frequency domain feature vector, the time domain feature vector and the energy feature vector of the target detection point respectively;
[0040] determine the target fusion feature set based on a comparison result of the preset similarity threshold and the cosine similarity;
[0041] perform feature fusion processing on the target fusion feature set to generate a fusion feature vector of the target detection point.
[0042] In a second aspect, the application provides a device for detecting insulation performance of a power battery cold plate, comprising:
[0043] a signal acquisition unit configured to acquire an ultrasonic echo signal of a target detection point on a surface of the power battery cold plate;
[0044] a graph generation unit configured to generate a frequency domain feature graph, a time domain feature graph and a time-frequency energy feature graph of the target detection point based on the ultrasonic echo signal of the target detection point;
[0045] a vector extraction unit configured to extract deep features corresponding to the frequency domain feature graph, the time domain feature graph and the time-frequency energy feature graph of the target detection point based on a graph feature extraction network, to obtain a frequency domain feature vector, a time domain feature vector and an energy feature vector of the target detection point;
[0046] a vector determination unit configured to perform feature fusion processing on the frequency domain feature vector, the time domain feature vector and the energy feature vector of the target detection point to generate a fusion feature vector of the target detection point;
[0047] a local output unit configured to input the fusion feature vector of the target detection point into an insulation layer classification network and output an insulation layer detection result of the target detection point;
[0048] a global output unit configured to determine an insulation performance detection result of the power battery cold plate based on the insulation layer detection result of the target detection point.
[0049] In a third aspect, an electronic device comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor is configured to implement the steps of the method for detecting insulation performance of a power battery cold plate according to any one of the first aspect when executing the computer program stored in the memory.
[0050] In a fourth aspect, the application provides a computer-readable storage medium storing a computer program, wherein the computer program is executable by a processor to implement the method for detecting insulation performance of a power battery cold plate according to any one of the first aspect.
[0051] In summary, the application realizes multi-dimensional feature analysis and intelligent detection of the insulation layer state by acquiring the ultrasonic echo signal of the target detection point on the surface of the power battery cold plate and generating a frequency domain, time domain and time-frequency energy feature map, and combining a feature map feature extraction network and an insulation layer classification network. The application effectively integrates signal features of different dimensions through multi-modal feature fusion, enhances the feature representation capability, can accurately capture subtle differences in the thickness and uniformity of the insulation layer, and compared with the traditional manual detection method, improves the accuracy and efficiency of the insulation layer detection, provides a more reliable technical solution for the insulation performance evaluation of the power battery cold plate, and ensures the safe and stable operation of the battery system. BRIEF DESCRIPTION OF DRAWINGS
[0052] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The accompanying drawings are included to provide a description of the preferred embodiments and are not meant to limit the present description. Moreover, the same reference numerals in the attached drawings indicate the same or similar components. In the drawings:
[0053] Figure 1 A detection method flow diagram of the insulation performance of the power battery cold plate provided by the embodiment of the application;
[0054] Figure 2 A detection device structure diagram of the insulation performance of the power battery cold plate provided by the embodiment of the application;
[0055] Figure 3 A detection electronic equipment structure diagram of the insulation performance of the power battery cold plate provided by the embodiment of the application. DETAILED DESCRIPTION
[0056] The terms "first", "second", "third", "fourth" and the like in the description, claims and drawings of the present application (if any) are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the use of these terms herein is to be construed as interchangeable in order to distinguish between the similar objects. It is also to be understood that the description of the embodiments of the present application is just one of the many possible specific embodiments which can be constructed in accordance with the present application. And, therefore, the description of specific embodiments of the present application is not intended to limit the scope of the present application, but it is constructed to provide the practical teaching with sufficiently sufficient to, together with the drawings, enable others skilled in the art to utilize the present application. Moreover, the term "comprising" and "having" and any variations thereof are intended to cover a non-exclusive inclusion, for example, a process, method, system, product or apparatus that comprises a list of steps or units is not necessarily limited to those steps or units that are clearly listed, but can include other steps or units that are not clearly listed or inherent to such processes, methods, products or apparatus. The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all embodiments.
[0057] Please refer toFigure 1 A flowchart of a detection method for the insulation performance of a power battery cold plate according to an embodiment of the present application is provided, which can specifically include the following steps:
[0058] S110, acquiring an ultrasonic echo signal of a target detection point on the surface of the power battery cold plate;
[0059] For example, the ultrasonic detection of the insulation layer of the power battery cold plate is based on the propagation and reflection characteristics of ultrasonic waves in different media. When the ultrasonic sensor emits a high-frequency pulse to the surface of the cold plate, the sound wave penetrates the insulation layer and is reflected at the interface between the insulation layer and the cold plate base. The time delay, intensity and waveform characteristics of the echo signal are directly related to the thickness and uniformity of the insulation layer. If the thickness of the insulation layer is uniform, the reflection law of the echo signal is stable. If there is a thickness abnormality or defect, the change in acoustic impedance of the sound wave propagation path will cause changes in the time delay and energy distribution of the echo signal, providing a physical basis for the quantitative analysis of the state of the insulation layer.
[0060] According to an embodiment of the present application, N detection points (N is an integer greater than 1) are arranged on the surface of the cold plate. An array sensor is used to synchronously acquire ultrasonic echo signals of each point, forming a detection network covering the surface of the cold plate. Compared with the traditional manual single-point detection, the multi-detection-point acquisition mode can comprehensively capture the local abnormality and overall uniformity information of the insulation layer through a set of spatially distributed signal characteristics. The echo signals of each detection point will be subsequently converted into frequency domain, time domain and energy feature maps. This multi-dimensional signal conversion mechanism maps the physical layer sound wave propagation characteristics into calculable digital features, laying a data foundation for subsequent intelligent analysis.
[0061] S120, generating a frequency domain feature map, a time domain feature map and a time-frequency energy feature map of the target detection point based on the ultrasonic echo signal of the target detection point;
[0062] For example, the ultrasonic echo signal contains multi-dimensional information of the thickness and uniformity of the insulation layer. The essence of generating the frequency domain, time domain and time-frequency energy feature maps is to map the one-dimensional sound wave signal into a multi-dimensional visual feature space. The frequency domain feature map reveals the frequency component distribution of the signal through Fourier transform. The abnormal thickness of the insulation layer will cause the energy distribution of a specific frequency band to shift. The time domain feature map is based on the time series dynamics of the signal amplitude. The waveform change is converted into a symmetric point distribution through polar coordinate mapping, reflecting the time delay characteristics of the insulation layer interface reflection. The time-frequency energy feature map combines short-time Fourier transform and filter banks to capture the energy distribution law of the signal in the time and frequency dimensions. The insulation layer defect will cause the energy of a specific frequency band to attenuate or mutate.
[0063] Firstly, the signal noise is eliminated by normalization, windowing and other pretreatments, and then the original acoustic wave signal is converted into a two-dimensional feature matrix by using frequency domain transform (such as Chirp-Z transform), time domain polar coordinate mapping (such as SDP method) and time-frequency analysis (such as short-time Fourier transform) respectively. Finally, the visual atlas is generated by color mapping, gradient calculation and other operations. This multi-dimensional feature conversion mechanism can comprehensively capture the subtle differences of the insulation layer thickness and uniformity from the three aspects of frequency component, time dynamic and energy distribution, and provides a rich data basis for subsequent deep feature extraction.
[0064] In S130, the frequency domain feature atlas, the time domain feature atlas and the time-frequency energy feature atlas of the target detection point are extracted by using the atlas feature extraction network, and the frequency domain feature vector, the time domain feature vector and the energy feature vector of the target detection point are obtained.
[0065] For example, the atlas feature extraction network is essentially a multi-layer feature abstraction mechanism based on convolutional neural network (CNN). Through the cascade operation of convolutional layer and pooling layer, the discriminative deep features are automatically extracted from the frequency domain, time domain and time-frequency energy feature atlas. For each type of feature atlas, the CNN slides in the spatial dimension through the learnable convolution kernel, captures the local patterns in the atlas (such as frequency distribution features of frequency domain atlas, symmetry point distribution rules of time domain atlas), and reduces the feature dimension through the pooling operation to suppress noise interference, so as to gradually convert the two-dimensional atlas information into high-dimensional abstract features.
[0066] After the multi-layer convolution and pooling processing, the atlas feature extraction network maps the three types of feature atlas into corresponding frequency domain, time domain and energy feature vectors. The convolutional layer is responsible for extracting the basic features such as edges and textures in the atlas, and the deep network further integrates these basic features to form semantic features related to the thickness and uniformity of the insulation layer. Finally, the features are compressed into fixed-dimensional vectors through full connection layer or specific convolution operation. These feature vectors not only retain the core information of the original atlas, but also have numerical representation form suitable for subsequent fusion processing, laying data foundation for multi-modal feature fusion.
[0067] In S140, the frequency domain feature vector, the time domain feature vector and the energy feature vector of the target detection point are fused to generate a fusion feature vector of the target detection point.
[0068] Exemplarily, the feature fusion processing integrates the frequency domain, time domain and energy feature vectors into a unified representation space through channel dimension splicing, so that the features of different physical dimensions interact (such as the correlation of the frequency domain frequency distribution feature and the time domain waveform dynamic feature). This operation retains the original discriminative information of each modal feature, while building a cross-domain feature correlation basis. Subsequently, the spliced features are cross-channel weighted through a 1x1 convolution layer, which automatically learns the contribution weights between different modalities, suppresses redundant information and strengthens key features, forming a preliminary fused convolution feature vector.
[0069] Based on the self-attention mechanism, dynamic weight distribution is performed on the convolution feature vector, and by calculating the correlation (Query-Key matching degree) between feature positions, higher weights are given to features that are strongly related to the insulating layer state. This process enables the model to focus on sensitive areas with abnormal thickness or uniformity changes (such as frequency energy mutation segments and time domain reflection delay points). Finally, the weighted features are added to the original convolution features through a residual connection, injecting discriminative information refined by attention while retaining the basic physical features, avoiding feature degradation problems in deep network training, and generating the final fused feature vector.
[0070] S150, input the fusion feature vector of the target detection point into the insulating layer classification network, and output the insulating layer detection result of the target detection point;
[0071] Exemplarily, the insulating layer classification network maps the fusion feature vector to the classification result of the insulating layer state through multi-level feature processing and attention mechanism. This network uses point convolution layer to expand feature channels, captures the correlation of local features through window attention backbone block, combines down-sampling layer to realize step-by-step compression of feature resolution, so that the network can extract key features of the insulating layer at different scales, and finally complete the mapping from feature space to classification space through global average pooling and fully connected layer, realize the quantitative evaluation of the thickness and uniformity of the insulating layer.
[0072] The window multi-head self-attention operation dynamically adjusts the weights of features in different regions, enhancing the sensitivity to insulating layer abnormalities; the down-sampling layer gradually integrates global context information to avoid local noise interference; the global average pooling compresses the spatial features into vector representation, and the fully connected layer maps it to the probability distribution of the insulating layer state through nonlinear transformation, and finally determines whether the insulating layer meets the design requirements based on the probability distribution, and outputs the detection result. This process realizes the intelligent conversion from multi-modal fusion features to insulating performance evaluation.
[0073] S160, based on the insulating layer detection result of the target detection point, determine the insulating performance detection result of the power battery cold plate.
[0074] Exemplarily, the insulation layer detection result of the power battery cold plate needs to be globally integrated based on the local results of N detection points, and through statistical analysis and threshold determination of the spatial distribution characteristics, the insulation performance evaluation from point to plane is realized. Since the result of a single detection point only reflects the local insulation layer state, and the uniformity and overall thickness of the insulation layer on the cold plate surface are the key to evaluation, the discrete local detection results need to be converted into global judgment through aggregation strategy to ensure that the evaluation covers the insulation layer integrity of the cold plate surface.
[0075] Based on the physical state (thickness, uniformity) of the insulation layer, the insulation performance is inferred, and insufficient thickness will lead to a decrease in insulation resistance and an increase in breakdown risk, while poor uniformity will cause inconsistent mechanical strength, indirectly affecting the stability of electrical insulation. By comparing the detected physical parameters (thickness mean, standard deviation, uniform point proportion) with the design threshold, the electrical performance of the insulation layer can be quantitatively evaluated to determine whether it meets the safety requirements of the power battery system.
[0076] In summary, in the embodiments of the present application, the ultrasonic echo signal of the target detection point on the surface of the power battery cold plate is obtained, the single ultrasonic echo signal is converted into frequency spectrum distribution, time domain waveform mapping and time-frequency energy heat map Figure 3 characteristic spectrum, which comprehensively captures the thickness abnormalities and uniformity changes of the insulation layer from the frequency component, time dynamics and energy distribution dimensions. For example, the frequency domain characteristic spectrum enhances the target frequency band resolution through Chirp-Z transformation, the time domain characteristic spectrum converts the waveform dynamics into a two-dimensional space pattern using polar coordinate symmetry mapping, and the time-frequency energy spectrum extracts the sound energy distribution rule through a filter bank. The three complement each other to form a complete characterization of the physical properties of the insulation layer. With the help of the spectrum feature extraction network, deep features are extracted and fused to generate a fusion feature vector. Then, the local detection result is output by the insulation layer classification network, and finally the overall insulation state is determined based on the multi-point detection result, realizing multi-dimensional feature analysis and intelligent detection of the insulation layer state. This method effectively integrates signal features of different dimensions through multi-modal feature fusion, enhances the feature representation capability, and accurately captures the subtle differences in insulation layer thickness and uniformity. Compared with traditional manual detection methods, the accuracy and efficiency of insulation layer detection are improved, the subjectivity and one-sidedness of manual operation are avoided, and a technical solution is provided for the insulation performance evaluation of the power battery cold plate, ensuring that the battery system avoids safety hazards such as current leakage and circuit short circuit caused by insulation layer problems during operation, and ensuring the safe and stable operation of the battery system.
[0077] In some examples, based on the ultrasonic echo signal of the target detection point, a frequency domain characteristic spectrum of the target detection point is generated, including:
[0078] The ultrasonic echo signal of the target detection point is normalized to generate a normalized signal of the target detection point;
[0079] perform windowed Fourier transform on the normalized signal of the target detection point based on a preset window function to generate an initial frequency spectrum of the target detection point;
[0080] perform frequency spectrum refinement on the target frequency band based on the initial frequency spectrum of the target detection point and the target frequency band to generate an enhanced frequency spectrum of the target detection point;
[0081] calculate a logarithmic amplitude value and a phase difference value of the enhanced frequency spectrum of the target detection point;
[0082] construct a frequency domain feature matrix of the target detection point based on the logarithmic amplitude value, the phase difference value, and a preset frequency distance attenuation factor;
[0083] perform quantile normalization and gradient amplitude calculation on the frequency domain feature matrix to generate quantile normalization results and gradient amplitude calculation results of the target detection point;
[0084] generate a frequency domain feature map of the target detection point based on the quantile normalization results and the gradient amplitude calculation results.
[0085] For example, when performing dynamic range normalization processing on the ultrasonic echo signal of the target detection point, first, the signal sequence is extracted and the absolute maximum max|X| (i.e., the maximum value of the absolute value of the sequence sample) is calculated; based on a preset simulation parameter β (β is an adjustment coefficient greater than 0, used to adjust the coordination amount ratio), the coordination adjustment amount ε (ε = β · max|X|, to avoid denominator anomaly when max|X| tends to zero, and to suppress noise amplification when the amplitude is close to zero) is derived; then, for each sample x i of the ultrasonic echo signal, the mapping is completed through the formula x i / (max|X|+ε) to realize signal normalization. The above method eliminates the direct current bias and retains the signal trend, constructs a normalization scale based on the absolute maximum value combined with the coordination amount, unifies the amplitude dynamic range of different detection points, provides stable and comparable input for subsequent frequency domain feature extraction, and adapts to the training and inference needs of the recognition model.
[0086] perform windowed Fourier transform on the normalized signal based on a preset window function (such as the Hanning window, the Blackman window, etc.), the purpose of which is to reduce spectral leakage and improve the accuracy of frequency domain analysis through the spectral shaping characteristics of the window function. The preset parameters (such as window length, overlap rate) of the window function are pre-set according to the frequency range and resolution requirements of the insulation layer detection, the time domain signal is divided into frames and multiplied by the window function, so that the signal satisfies the periodicity assumption in the time domain, and then the time domain signal is converted into a frequency domain representation through Fourier transform to generate an initial frequency spectrum containing the energy distribution of each frequency component. This process converts the physical characteristics related to the thickness and uniformity of the insulation layer into frequency energy distribution characteristics, providing original data support for subsequent frequency spectrum refinement.
[0087] The frequency spectrum refinement based on the initial frequency spectrum and the preset target frequency band (i.e., the frequency interval sensitive to the thickness change of the insulation layer) is a key operation for accurately capturing the characteristic frequency of the insulation layer by improving the frequency resolution of the target frequency band. Specifically, this step uses algorithms such as Chirp Z transform or frequency domain interpolation to locally amplify the frequency spectrum of the target frequency band, increasing the number of sampling points within a unit frequency interval, so that the characteristic frequency components blurred in the initial frequency spectrum are finely characterized. For example, if the frequency resolution of the initial frequency spectrum is 10 Hz, the target frequency band of 1000-2000 Hz is refined by 10 times, and the resolution can be improved to 1 Hz, so that the characteristic frequency shift or energy peak value change caused by the thickness abnormality of the insulation layer can be more accurately identified.
[0088] The calculation of the logarithmic amplitude value and the phase difference value of the enhanced frequency spectrum is to convert the frequency energy distribution into a numerical representation suitable for the construction of the feature matrix. The calculation of the logarithmic amplitude value compresses the dynamic range of the frequency spectrum, making the difference between strong signals and weak signals more easily handled numerically, while preserving the relative energy relationship of the frequency spectrum; the calculation of the phase difference value reflects the phase change trend of adjacent frequency points, which is closely related to the acoustic characteristics of the insulation layer interface. When the thickness of the insulation layer is uniform, the phase change has regularity, and the thickness abnormality will cause phase mutation. The calculation of these two parameters provides multi-dimensional features with both energy distribution and phase characteristics for the subsequent construction of the frequency domain feature matrix.
[0089] The frequency domain feature matrix is constructed based on the logarithmic amplitude value, the phase difference value, and the preset frequency distance attenuation factor, which weights the feature importance of different frequency points through the attenuation factor. The frequency distance attenuation factor is preset as a function (such as an exponential decay function) that decreases with the increase of the distance between the frequency and the center frequency, the purpose of which is to suppress the interference of distant frequencies and highlight the characteristic frequency components related to the thickness of the insulation layer. In the specific construction process, the frequency is taken as the horizontal axis, and the logarithmic amplitude value and the phase difference value are taken as the vertical axis. The weighted logarithmic amplitude value and the phase difference value of each frequency point are taken as the matrix elements to form a two-dimensional frequency domain feature matrix. This matrix not only preserves the energy distribution and phase information of the frequency spectrum, but also enhances the discriminability of the features through the attenuation factor, providing structured data for subsequent graph generation.
[0090] The quantile normalization and gradient amplitude calculation of the frequency domain feature matrix are to eliminate the influence of abnormal values in the matrix and extract the local change features of the frequency spectrum. Quantile normalization maps the matrix elements to a preset quantile range (10% quantile and 90% quantile), so that the feature matrices of different detection points have a uniform numerical distribution, enhancing the feature contrast; the gradient amplitude calculation calculates the partial derivatives of the matrix in the row direction and the column direction, and calculates the Euclidean distance to obtain the intensity information of the spatial change of the matrix. This step strengthens the edge features of the thickness abnormality area and improves the visual recognition.
[0091] The frequency domain feature atlas is generated based on the quantile normalization result and the gradient amplitude calculation result, the numerical feature matrix is converted into a visual two-dimensional atlas to adapt to the input requirements of the deep learning network. In specific implementation, the normalized logarithmic amplitude value and the gradient amplitude value are respectively mapped into the brightness and hue of the color through color mapping (such as a heat map), to form a visual atlas that fuses the energy distribution and the change feature. Each pixel point in the atlas corresponds to a frequency point, and the color feature directly reflects the energy intensity and the change gradient of the frequency point, so that the frequency domain features (such as the energy peak shift and the gradient mutation region) caused by the insulation layer thickness abnormality are presented in a visual form, to provide clear visual input features for the subsequent atlas feature extraction network.
[0092] In summary, the embodiment of the present application constructs a frequency domain visual atlas containing the insulation layer thickness and uniformity sensitive features through multiple-step processing such as normalization processing, windowed Fourier transform, target frequency band refinement, logarithmic amplitude and phase calculation, attenuation factor weighting, quantile normalization and gradient amplitude extraction. Compared with the traditional frequency domain analysis means, the embodiment of the present application improves the resolution of the feature frequency on the one hand through target frequency band refinement, and can capture subtle frequency shifts that are difficult to identify by traditional methods; on the other hand, through attenuation factor weighting and gradient amplitude calculation, the discriminability of the features is enhanced, so that the atlas can more accurately represent the insulation layer state. In addition, quantile normalization ensures the stability of the atlas features, provides standardized input for the subsequent deep learning network feature extraction, and finally realizes the dual improvement of the insulation layer detection precision and efficiency.
[0093] In some examples, based on the ultrasonic echo signal of the target detection point, a time domain feature atlas of the target detection point is generated, including:
[0094] Based on the ultrasonic echo signal of the target detection point, the signal maximum value, the signal minimum value and the current time point amplitude of the target detection point are determined;
[0095] Based on the preset lag factor, the lag time point amplitude of the target detection point is determined;
[0096] Based on the signal maximum value and the signal minimum value, the current time point amplitude of the target detection point is normalized and calculated to generate the normalized polar radius of the target detection point;
[0097] Based on the signal maximum value and the signal minimum value, the lag time point amplitude of the target detection point is normalized and calculated to generate the lag time point amplitude normalization quantity of the target detection point;
[0098] The amplitude of the target detection point at the lag time point is calculated based on the normalized amplitude at the lag time point, a preset reference angle, and a preset angle gain factor, to generate a first deflection angle and a second deflection angle of the target detection point, wherein the first deflection angle = the preset reference angle + (the normalized amplitude at the lag time point x the preset angle gain factor), the second deflection angle = the preset reference angle - (the normalized amplitude at the lag time point x the preset angle gain factor), and the normalized amplitude at the lag time point = (the amplitude at the lag time point - the minimum value of the signal) / (the maximum value of the signal - the minimum value of the signal);
[0099] Based on the normalized polar radius, the first deflection angle, and the second deflection angle, a polar coordinate mapping point set of the target detection point is constructed.
[0100] Based on the polar coordinate mapping point set of the target detection point, a symmetric point distribution map of the target detection point is generated, and the symmetric point distribution map is taken as a time domain feature map of the target detection point.
[0101] For example, based on the ultrasonic echo signal of the target detection point, the maximum value, the minimum value, and the amplitude at the current time point of the signal are first determined. The maximum value and the minimum value of the signal reflect the overall amplitude range of the echo signal, which is the basis for subsequent normalization processing; the amplitude at the current time point represents the instantaneous intensity of the signal at a certain time. The acquisition of these three parameters is realized by traversing the time domain sequence of the echo signal, which provides numerical support for the subsequent normalization of the amplitude at the lag time point, the calculation of the polar radius, and the derivation of the deflection angle, ensures the accuracy and consistency of the time domain feature extraction, and avoids the feature deviation caused by the fluctuation of the signal amplitude.
[0102] The preset lag factor is an empirical parameter (for example, set to 5 or 10 sampling points) that is set in advance according to the time resolution requirement of the power battery cold plate insulation layer detection, and is used to determine the time interval between the current time point and the lag time point. Based on the factor, the amplitude of the nth sampling point (n is the value of the lag factor) after the current time point in the echo signal sequence is selected as the amplitude at the lag time point. Through this operation, the dynamic change feature of the signal in the time dimension is captured. When the insulation layer thickness is uniform, the amplitudes at the current and lag time points have strong correlation; if there is a thickness abnormality, the time correlation of the amplitudes will change due to the change of the sound wave reflection path, and through the comparison of the lag amplitude and the current amplitude, the time domain dynamic feature related to the state of the insulation layer can be effectively extracted.
[0103] Based on the maximum and minimum values of the signal, the amplitude at the current time point is normalized to calculate the amplitude mapping to the interval [0, 1], eliminating the influence of the signal amplitude difference at different detection points. The difference between the amplitude at the current time point and the minimum value of the signal is divided by the difference between the maximum and minimum values to obtain the normalized result as the normalized polar radius. In the polar coordinate representation system, the normalized value corresponds to the radius parameter in the polar coordinate, and its physical meaning is to convert the relative intensity of the signal amplitude into the polar radius length. The larger the amplitude, the longer the polar radius, and vice versa. The geometric meaning is the distance from the point to the origin in the polar coordinate system, and its numerical value represents the relative level of the sound wave reflection intensity at the current time, eliminating the interference caused by the absolute amplitude fluctuation of the signal. This conversion makes the amplitude characteristics of the time domain signal visualized in the form of polar coordinates, laying a foundation for the construction of the polar coordinate mapping point set.
[0104] Based on the maximum and minimum values of the signal, the amplitude at the current time point is normalized to calculate the amplitude mapping to the interval [0, 1], eliminating the influence of the signal amplitude difference at different detection points. The difference between the amplitude at the current time point and the minimum value of the signal is divided by the difference between the maximum and minimum values to obtain the normalized result as the normalized polar radius. In the polar coordinate representation system, the normalized value corresponds to the radius parameter in the polar coordinate, and its physical meaning is to convert the relative intensity of the signal amplitude into the polar radius length. The larger the amplitude, the longer the polar radius, and vice versa. The geometric meaning is the distance from the point to the origin in the polar coordinate system, and its numerical value represents the relative level of the sound wave reflection intensity at the current time, eliminating the interference caused by the absolute amplitude fluctuation of the signal. This conversion makes the amplitude characteristics of the time domain signal visualized in the form of polar coordinates, laying a foundation for the construction of the polar coordinate mapping point set.
[0105] Based on the normalized value of the amplitude at the lag time point, the preset reference angle and the preset angle gain factor, the symmetric deflection calculation is performed, aiming to convert the change of the amplitude at the lag time point into the angle feature in the polar coordinate, forming an angle distribution with symmetry. The preset reference angle is a reference angle in the polar coordinate, used to determine the center position of the deflection; the preset angle gain factor is used to adjust the sensitivity of the angle deflection, and its numerical value is set in advance according to the accuracy requirement of the insulation layer thickness detection. The normalized value of the amplitude at the lag time point is obtained by calculating the difference between the maximum and minimum values, reflecting the relative position of the amplitude in the overall signal. Through the symmetric calculation of the reference angle ± (normalized value x gain factor), the first deflection angle and the second deflection angle are generated, so that the change of the lag amplitude is represented in the form of angle symmetry. When the lag amplitude is consistent with the current amplitude, the deflection angles are symmetrically distributed; when there is a difference in amplitude, the angle symmetry is broken, and this symmetry change is directly related to the uniformity of the insulation layer thickness.
[0106] The normalized polar radius is combined with the first deflection angle and the second deflection angle to construct a polar coordinate mapping point set, and the amplitude and time correlation characteristics of the time domain signal are converted into point distribution in the polar coordinate space. Each point is composed of (normalized polar radius, first deflection angle) and (normalized polar radius, second deflection angle), which correspond to two symmetric points in the polar coordinate respectively. This point set construction method utilizes the decoupling characteristics of the polar coordinate system for amplitude and angle. The polar radius represents the relative strength of the signal amplitude, and the angle represents the correlation change of the signal in the time dimension. By converting the amplitude and lag amplitude of each time point into symmetric polar coordinate points, a point set is formed that can reflect the time domain dynamic characteristics of the signal, providing structured data for subsequent generation of visual atlas.
[0107] When generating a symmetric point distribution graph based on the polar coordinate mapping point set, the polar coordinate points are first converted into Cartesian coordinates (x, y), where x = normalized polar radius x cos(angle), y = normalized polar radius x sin(angle), and angle is the first deflection angle and the second deflection angle. Then, all symmetric points are plotted on a two-dimensional plane to form a point distribution atlas with a preset reference angle as the symmetry axis. In the atlas, the distribution density, symmetry, and degree of deviation from the reference axis of the points directly reflect the time domain characteristics of the ultrasonic echo signal. When the insulation layer thickness is uniform, the point distribution has good symmetry. If there is a thickness anomaly or defect, the symmetry of the point distribution will be destroyed, and local concentration or deviation will occur. This visual atlas converts abstract time domain signal characteristics into intuitive geometric distribution, facilitating subsequent extraction of key features related to the insulation layer state through a deep learning network.
[0108] In summary, in the embodiments of the present application, through multiple steps such as signal extreme value determination, lag amplitude extraction, normalized polar radius calculation, symmetric deflection angle derivation, polar coordinate point set construction, and symmetric point distribution graph generation, geometric characterization of ultrasonic echo signal time domain characteristics is achieved. Compared with traditional time domain analysis methods, on the one hand, through polar coordinate mapping and symmetric deflection calculation, the amplitude change and time correlation of the signal are converted into polar radius and angle characteristics with physical meaning, enhancing the interpretability of the time domain characteristics. On the other hand, through the parameterization design of the preset reference angle and the gain factor, the sensitivity to insulation layer thickness anomalies can be flexibly adjusted to adapt to the needs of different detection scenarios. In addition, the generated symmetric point distribution graph can directly reflect the time domain dynamic change of the signal, providing a visual input containing insulation layer uniformity information for the subsequent atlas feature extraction network, effectively improving the extraction accuracy of time domain features in insulation layer detection.
[0109] In some examples, based on the ultrasonic echo signal of the target detection point, a time-frequency energy feature atlas of the target detection point is generated, including:
[0110] perform time-frequency transformation on the ultrasonic echo signal of the target detection point to generate a time-frequency spectrum energy matrix of the target detection point;
[0111] perform frequency band decomposition on the time-frequency spectrum energy matrix of the target detection point through a preset filter bank to extract an energy distribution feature of the target detection point;
[0112] construct a time-frequency energy distribution matrix of the target detection point based on the energy distribution feature of the target detection point;
[0113] convert the time-frequency energy distribution matrix of the target detection point into a visualized thermal map, and use the visualized thermal map as a time-frequency energy feature map of the target detection point.
[0114] Exemplarily, time-frequency transformation is performed on the ultrasonic echo signal of the target detection point to generate a time-frequency spectrum energy matrix representing the time-space distribution of signal energy. Specifically, short-time Fourier transform (STFT) is used as the core algorithm, the time-domain signal is segmented into a fixed-length frame sequence through a preset window function (such as a Hanning window), Fourier transform is performed on each frame of signal, the energy amplitude of each time point at different frequencies is calculated, and a two-dimensional time-frequency spectrum energy matrix is formed. The horizontal axis of the matrix represents the time sequence, the vertical axis represents the frequency component, and the matrix element value represents the signal energy intensity at a specific time and frequency point. The window function length is set to 1-5 μs according to the ultrasonic wave period, and the frame shift amount is set to 1 / 4 to 1 / 2 of the frame length, so as to balance the time resolution and calculation efficiency, and ensure that the micro-time delay feature caused by the change of the insulation layer thickness can be accurately captured.
[0115] The time-frequency spectrum energy matrix is decomposed through a preset filter bank to divide the wide-band time-frequency energy distribution into a plurality of narrow-band sub-band energy features, so as to capture the feature band energy change related to the insulation layer thickness. The filter bank adopts a Gammatone band-pass filter design, and the center frequency and bandwidth are pre-set according to the acoustic characteristics of the insulation layer material and the thickness detection range. During the decomposition process, each filter performs convolution operation on the time-frequency spectrum energy matrix to extract the energy distribution of the corresponding frequency band and form a feature sequence of the energy change with time of each frequency band. This frequency band decomposition operation can effectively suppress the interference of irrelevant frequency bands and highlight the feature energy distribution related to the insulation layer state, for example, when the insulation layer thickness is abnormal, the energy of a specific frequency band will appear attenuation or mutation.
[0116] Based on the energy distribution characteristics of each frequency band, the time-frequency energy distribution matrix is constructed, which is the process of integrating the multi-dimensional energy features obtained by frequency band decomposition into structured data. The matrix takes time as the horizontal axis and the preset frequency band as the vertical axis, and each matrix element corresponds to the energy value of a certain time point in a specific frequency band. When constructing, the energy features of each frequency band need to be standardized (such as Z-score standardization) first, and then arranged in time sequence and frequency band order to form a two-dimensional matrix. This matrix structure can not only retain the energy distribution details in time-frequency domain, but also reflect the dynamic characteristics of insulation layer thickness and uniformity through the numerical changes of matrix elements, for example, when there is a local thickness deficiency in the insulation layer, the energy of a specific frequency band at the corresponding time point will decrease significantly, which is manifested as a local low-energy value area in the matrix.
[0117] Converting the time-frequency energy distribution matrix into a visual heat map is a key step of converting the numerical matrix into an intuitive visual representation through color mapping. In specific implementation, a preset color mapping rule (such as Jet, Hot, etc.) is used to map the energy values in the matrix to different colors after normalization, with high energy values corresponding to warm colors (such as red) and low energy values corresponding to cold colors (such as blue), forming a two-dimensional visual map of time, frequency and energy three-dimensional information. The color distribution and gradient change in the heat map directly reflect the time-frequency energy features of the ultrasonic echo signal, and the energy distribution presents regular changes when the insulation layer is uniform; when there is a defect, color mutation or abnormal color block will appear in the map. This visualization process not only facilitates manual interpretation, but also provides an image-based input suitable for feature extraction for subsequent deep learning networks.
[0118] In summary, the embodiments of the present application realize the extraction and representation of time-frequency energy features in the ultrasonic echo signal through multiple-step processing such as time-frequency transformation, frequency band decomposition, matrix construction and heat map visualization. Compared with traditional energy analysis methods, on the one hand, through the frequency band decomposition of the preset filter set, the feature frequency band energy related to the insulation layer thickness is extracted, the interference of environmental noise and irrelevant frequency bands is effectively suppressed, and the discriminability of the energy features is improved; on the other hand, through the visualization processing of the heat map, the abstract energy distribution is converted into intuitive color features, which not only facilitates manual rapid positioning of abnormal areas, but also provides a high-quality input containing time and space energy information for deep learning models.
[0119] In some examples, the atlas feature extraction network includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a third convolutional layer, a second pooling layer, a fourth convolutional layer, a third pooling layer, and a fifth convolutional layer. Based on the atlas feature extraction network, a deep feature of a frequency domain feature atlas of a target detection point is extracted to obtain a frequency domain feature vector of the target detection point, including:
[0120] Based on the frequency domain feature atlas of the target detection point, a first input image of the input layer is determined.
[0121] The first input image is subjected to a first convolution operation by a first convolution layer to generate a first feature map of the target detection point; the first feature map is subjected to a first pooling operation by a first pooling layer to generate a second feature map of the target detection point; the second feature map is subjected to a second convolution operation by a second convolution layer to generate a third feature map of the target detection point; the third feature map is subjected to a third convolution operation by a third convolution layer to generate a fourth feature map of the target detection point; the fourth feature map is subjected to a second pooling operation by a second pooling layer to generate a fifth feature map of the target detection point; the fifth feature map is subjected to a fourth convolution operation by a fourth convolution layer to generate a sixth feature map of the target detection point; the sixth feature map is subjected to a third pooling operation by a third pooling layer to generate a seventh feature map of the target detection point; and the seventh feature map is subjected to a fifth convolution operation by a fifth convolution layer to generate a frequency domain feature vector of the target detection point.
[0122] For example, when determining the first input image of the input layer based on the frequency domain feature map of the target detection point, the frequency domain feature map needs to be preprocessed to adapt to the size and data format requirements of the input layer. Specifically, the image size of the preset input layer is 224x224 pixels, and the RGB three-channel format is adopted. If the original size of the frequency domain feature map does not conform, scaling is performed by methods such as bilinear interpolation. At the same time, the pixel values of the map are normalized to the interval [0, 1] to eliminate the difference in numerical scale. The essence of this step is to convert the visual frequency domain feature map into a standard input format that can be processed by the neural network, ensuring that the subsequent convolution operation can effectively extract features and avoiding feature extraction bias caused by inconsistent input formats.
[0123] The first convolution layer uses a 7x7 size convolution kernel, sets the stride to 2, and fills 3 pixels. The convolution operation is performed on the input 224x224 image, and a 112x112 feature map of 64 channels is output. This operation captures the global pattern of the image (such as the overall trend of the frequency domain energy distribution) by expanding the size of the convolution kernel, and sets the stride to reduce the spatial resolution while preserving key information, laying the foundation for feature extraction in subsequent layers.
[0124] The first pooling layer uses a 3x3 pooling window, sets the stride to 2, and fills 1 pixel. The maximum pooling operation is performed on the 112x112 feature map output by the first convolution layer, and the maximum response value of each local region is extracted, outputting a 56x56 feature map of 64 channels. This operation reduces the spatial dimension of the feature map, reduces the amount of calculation, and suppresses noise interference. The pooling operation preserves the maximum value in the local region, making the feature representation translationally invariant, for example, the frequency domain feature shift caused by the thickness abnormality of the insulating layer can still be effectively captured after pooling. The generated second feature map is halved in size, but retains the key frequency domain feature information, providing a reduced input for subsequent deep convolution to extract more abstract features.
[0125] The second convolutional layer uses a 3x3 convolutional kernel, with a stride of 1 and padding of 1 pixel. The 56x56 feature map output by the first pooling layer is subjected to convolutional operation, outputting a 56x56 feature map with 128 channels. This stage extracts local features (such as energy peak regions in frequency domain maps) through a small size convolutional kernel, and the padding operation keeps the feature map size unchanged, ensuring that detailed information is not lost.
[0126] The third convolutional layer also uses a 3x3 convolutional kernel, with a stride of 1 and padding of 1 pixel. The 56x56 feature map output by the second convolutional layer is subjected to convolutional operation, outputting a 56x56 feature map with 256 channels. This layer further deepens the feature abstraction capability, and through the increase of channel number, it strengthens the ability to capture complex patterns (such as frequency energy mutations or gradient edges).
[0127] The second pooling layer uses a 3x3 pooling window, with a stride of 2 and padding of 1 pixel. The 56x56 feature map output by the third convolutional layer is subjected to maximum pooling operation, outputting a 28x28 feature map with 256 channels. This operation again reduces the spatial dimension, gradually focusing on more discriminative high-level semantic features.
[0128] The fourth convolutional layer uses a 3x3 convolutional kernel, with a stride of 1 and padding of 1 pixel. The 28x28 feature map output by the second pooling layer is subjected to convolutional operation, outputting a 28x28 feature map with 512 channels. This layer integrates multi-scale features by expanding the channel dimension, improving the network's ability to distinguish subtle differences in frequency domain features.
[0129] The third pooling layer uses a 3x3 pooling window, with a stride of 2 and padding of 1 pixel. The 28x28 feature map output by the fourth convolutional layer is subjected to maximum pooling operation, outputting a 14x14 feature map with 512 channels. This step realizes the final dimension reduction of the feature map, providing high-density semantic information for feature vector generation.
[0130] The fifth convolutional layer uses a 1x1 convolutional kernel, with a stride of 1 and no padding operation. The 14x14 feature map output by the third pooling layer is subjected to convolutional operation, outputting a 1x14x14-dimensional feature vector. This operation realizes channel dimension compression through 1x1 convolution, fusing the 512-channel feature map into a single-channel feature vector, retaining key discriminative information while adapting to subsequent processing needs.
[0131] In summary, the embodiment of the present application realizes the layer-by-layer abstraction of the frequency domain feature atlas from the primary feature to the deep semantic feature through the input layer preprocessing, the cascade design of the multi-layer convolution and the pooling. The advantages of the network structure are as follows: firstly, the multi-layer convolution operation can capture the multi-scale features from the local edge to the global energy distribution in the frequency domain atlas through the convolution kernel and the channel configuration of different sizes, and accurately identify the frequency domain feature mode corresponding to the insulation layer thickness change; secondly, the progressive dimension reduction design of the pooling layer reduces the calculation amount while enhancing the anti-noise ability and the translation invariance of the feature, avoiding the misjudgment caused by the position deviation of the detection point; thirdly, the end-to-end training mechanism of the network can automatically optimize the parameters of each layer, so that the generated frequency domain feature vector has high discriminability and low redundancy, which improves the automation degree and the feature extraction precision of the insulation layer detection compared with the traditional manual feature extraction method, and provides a high-quality frequency domain feature input for multi-modal feature fusion and intelligent classification.
[0132] In some examples, the time domain feature vector and the energy feature vector are realized by the same method as the frequency domain feature vector, that is, based on the input layer, the convolution layers and the pooling layers of the atlas feature extraction network, the time domain feature atlas and the energy feature atlas are processed respectively: the time domain feature atlas and the energy feature atlas are determined as the input image of the input layer, the corresponding feature map is generated through the convolution operation of the first convolution layer, and then the time domain feature vector and the energy feature vector are finally extracted through the convolution and pooling operations of the second convolution layer, the third convolution layer, the second pooling layer, the fourth convolution layer, the third pooling layer and the fifth convolution layer in turn after the first pooling layer.
[0133] In some examples, the frequency domain feature vector, the time domain feature vector and the energy feature vector of the target detection point are subjected to feature fusion processing to generate a fusion feature vector of the target detection point, including:
[0134] The frequency domain feature vector, the time domain feature vector and the energy feature vector of the target detection point are subjected to normalization processing to generate a normalized frequency domain feature vector, a normalized time domain feature vector and a normalized energy feature vector of the target detection point;
[0135] Based on the normalized frequency domain feature vector, the normalized time domain feature vector and the normalized energy feature vector, a feature splicing operation is performed along the channel dimension to generate a spliced feature vector of the target detection point;
[0136] Based on the preset convolution kernel weight, a one-dimensional convolution operation is performed on the spliced feature vector to generate a convolution feature vector of the target detection point;
[0137] The convolution feature vector is subjected to a nonlinear transformation to generate an activation feature vector of the target detection point;
[0138] The self-attention weight matrix and the activation feature vector are subjected to a weighted summation operation to generate a weighted feature vector of the target detection point.
[0139] The self-attention weight matrix and the activation feature vector are subjected to a weighted summation operation to generate a weighted feature vector of the target detection point.
[0140] The weighted feature vector and the activation feature vector are subjected to a residual connection operation to generate a residual feature vector of the target detection point.
[0141] The residual feature vector is subjected to a nonlinear activation to generate a fusion feature vector of the target detection point.
[0142] The frequency domain feature vector, the time domain feature vector, and the energy feature vector of the target detection point are subjected to normalization processing to generate a normalized frequency domain feature vector, a normalized time domain feature vector, and a normalized energy feature vector of the target detection point. This step maps the three types of feature vectors to a unified numerical scale by calculating the mean and standard deviation of each feature vector using the zero-mean unit-variance standardization formula, eliminating the dimensional differences caused by different feature sources. Specifically, normalization processing ensures that the numerical distribution of frequency spectrum features, time waveform features, and time-frequency energy features is of the same order of magnitude, avoiding weight bias caused by feature amplitude differences in the subsequent fusion process and establishing a fair weighting basis for feature concatenation.
[0143] Based on the normalized frequency domain feature vector, the time domain feature vector, and the energy feature vector, a feature concatenation operation is performed along the channel dimension to generate a concatenated feature vector of the target detection point. The concatenation operation sequentially connects the three types of feature vectors in the channel direction. If the dimensions of each feature vector are 1x196 (14x14), the concatenated feature vector has a dimension of 3x196. This operation preserves the original spatial structure and discriminative information of each modal feature, while building a physical basis for cross-domain feature correlation, enabling frequency distribution, time domain dynamic changes, and time-frequency energy features to interact in a unified representation space, providing an integrated input containing multi-source information for subsequent convolution operations.
[0144] Based on the preset convolution kernel weight parameters, a one-dimensional convolution operation is performed on the concatenated feature vector to generate a convolution feature vector of the target detection point. The convolution kernel size is fixed at 1x1, the step is set to 1, and there is no padding operation. The concatenated features are weighted and fused across channels through a learnable weight matrix. This operation essentially performs a linear combination of the three types of feature vectors, automatically learning the contribution weights of frequency domain, time domain, and energy features, suppressing redundant information and enhancing key features. For example, when the insulation layer thickness anomaly is mainly reflected in the frequency domain feature change, the convolution kernel automatically increases the fusion weight of the frequency domain feature, and the output dimension is compressed to 1x196, achieving feature dimension reduction and information purification.
[0145] The nonlinear transformation is performed on the convolution feature vector to generate an activation feature vector of the target detection point using a ReLU activation function. The ReLU function serves to introduce nonlinear expression capability to the convolution feature, thereby enhancing the fitting capability of the model to complex feature patterns. This step maps the linear feature output by the convolution to a nonlinear space, so that the fusion process can capture nonlinear correlations of the insulation layer thickness and uniformity changes. For example, the truncation operation of the activation function on negative features can filter noise interference, retain positive activation features that are strongly correlated with the insulation layer state, and improve the robustness of feature discrimination.
[0146] Based on the query linear transformation matrix, the key linear transformation matrix, and the value linear transformation matrix, self-attention weight calculation is performed on the activation feature vector to generate a self-attention weight matrix of the target detection point. The three transformation matrices are obtained through model training, and the dimensions are all 196x196. First, the activation feature vector is multiplied by the query matrix and the key matrix respectively to generate a query vector and a key vector; then the query-key similarity matrix is calculated through dot product operation; finally, after scaling processing (divided by the square root of the dimension of the key vector) and Softmax normalization, a 196x196-dimensional self-attention weight matrix is output. The weight matrix quantifies the correlation between feature positions, so that the model focuses on the feature regions of the insulation layer that are sensitive to abnormalities (such as frequency energy mutation points or time domain reflection delay sections).
[0147] Weighted summation operation is performed on the self-attention weight matrix and the activation feature vector to generate a weighted feature vector of the target detection point. In specific implementation, the activation feature vector is first multiplied by the value linear transformation matrix to generate a value vector; then the self-attention weight matrix is multiplied by the value vector to obtain the weighted fusion result. This step assigns feature importance according to the weight matrix, for example, the high-weight feature region (weight value > 0.8) that is sensitive to the thickness change of the insulation layer is enhanced in the fusion result, while the low-weight region (weight value < 0.2) is weakened, thereby refining the feature representation with stronger discrimination and improving the detection sensitivity to subtle defects of the insulation layer.
[0148] Residual connection operation is performed on the weighted feature vector and the activation feature vector to generate a residual feature vector of the target detection point. This operation adds the weighted feature vector and the activation feature vector element by element to ensure that the basic physical features are not completely covered by the attention mechanism, while injecting the discriminative information refined by attention. For example, when the activation feature vector contains the basic frequency domain features of the insulation layer thickness, the weighted feature vector strengthens the abnormal mutation information therein, and the residual connection makes the two complementary to form a more comprehensive feature representation. This design effectively avoids the gradient vanishing problem in deep network training, and ensures the complete transmission of feature information.
[0149] The residual feature vector is nonlinearly activated, and a ReLU function is used to generate the final fusion feature vector of the target detection point. This activation operation further filters the negative noise that may occur after the residual connection, and outputs pure positive features. The fusion feature vector dimension remains 1x196, which not only contains the deep integration information of multi-modal features, but also has a structured format suitable for the processing of the insulation layer classification network, providing high-discriminative input features for subsequent classification decisions.
[0150] In summary, the embodiments of the present application eliminate scale differences through normalization, integrate multi-modal information through channel splicing, extract cross-domain correlations through one-dimensional convolution, enhance expression ability through nonlinear transformation, dynamically allocate weights through self-attention, retain basic features through residual connection, and optimize feature distribution through final nonlinear activation. The deep fusion of frequency domain, time domain and energy features is realized. The feature vector generated through this fusion processing can comprehensively represent the thickness uniformity and interface defects of the insulation layer, providing input features with comprehensiveness and high precision for the insulation layer classification network.
[0151] In some examples, the insulation layer classification network includes a point convolution layer, a first window attention backbone block, a first down-sampling layer, a second window attention backbone block, a second down-sampling layer, a third window attention backbone block, a global average pooling layer, and a fully connected layer. The fusion feature vector of the target detection point is input into the insulation layer classification network, and the insulation layer detection result of the target detection point is output, including:
[0152] The fusion feature vector of the target detection point is subjected to a channel expansion operation through the point convolution layer, generating an expanded feature map of the target detection point;
[0153] Based on a preset window division rule, the expanded feature map is divided into a plurality of non-overlapping windows;
[0154] The plurality of non-overlapping windows are subjected to window multi-head self-attention operation through the first window attention backbone block, generating a first window feature map of the target detection point;
[0155] The first window feature map is subjected to first down-sampling operation through the first down-sampling layer, generating a first down-sampled feature map of the target detection point;
[0156] The first down-sampled feature map is subjected to window multi-head self-attention operation through the second window attention backbone block, generating a second window feature map of the target detection point;
[0157] The second window feature map is subjected to second down-sampling operation through the second down-sampling layer, generating a second down-sampled feature map of the target detection point;
[0158] The second down-sampled feature map is subjected to window multi-head self-attention operation through the third window attention backbone block, generating a third window feature map of the target detection point;
[0159] The third window feature map is subjected to a spatial dimension compression operation by a global average pooling layer to generate a classification feature vector of the target detection point;
[0160] The classification feature vector is subjected to a dimension mapping operation by a fully connected layer to generate an insulating layer classification vector of the target detection point;
[0161] Based on the insulating layer classification vector, an insulating layer detection result of the target detection point is determined.
[0162] Exemplarily, a channel expansion operation is performed on the fusion feature vector of the target detection point by a point convolution layer, and the original 1x14x14-dimensional fusion feature vector is expanded to 128 channels by a 1x1 convolution kernel to generate an expanded feature map with a size of 128x14x14. This operation increases the feature channel dimension to adapt the embedding space requirement of the subsequent Transformer module while retaining the spatial resolution of the feature map, providing a structurally complete input data for window division. The preset channel expansion multiple (128 times) is determined by experimental optimization according to the complexity of the insulating layer features, ensuring a balance between feature expression ability and computational efficiency.
[0163] Based on a preset window division rule, the expanded feature map of 128x14x14 is divided into 4x4 non-overlapping windows, each with a fixed size of 7x7 pixels. The window division rule requires that the window size is an integer multiple of the spatial size of the feature map (14÷7=2), ensuring that the feature map is completely covered without overlapping areas. This step decomposes the global features into local window units, limiting the self-attention calculation range within the window and reducing the computational complexity. The preset window size (7x7) is set according to the physical size of the local defects of the insulating layer, so that each window corresponds to about 1-2 cm 2 area on the surface of the cold plate, matching the spatial distribution characteristics of the insulating layer thickness abnormalities.
[0164] A window multi-head self-attention operation is performed on the divided multiple non-overlapping windows by a first window attention backbone block. The backbone block is preset with 4 attention heads, each of which independently calculates the correlation weight between the 49 feature positions (7x7) within the window, capturing the local feature dependency relationship through a query-key value matching mechanism. A 4-pixel window displacement strategy is introduced in the calculation process to produce a partial overlapping area between adjacent windows, realizing cross-window information interaction. This step outputs a first window feature map with a size of 128x14x14, but strengthens the local feature relevance related to the insulating layer thickness within the window, such as enhancing the sensitivity to the abnormal reflection of sound waves caused by coating micro-cracks.
[0165] The first window feature map is subjected to a down-sampling operation by a first down-sampling layer. The operation first concatenates the features of adjacent 2x2 windows (a total of 4 windows) along the channel dimension, increasing the number of channels to 512 (128x4); then a 1x1 convolution kernel is used to compress the number of channels to 256, while reducing the spatial resolution from 14x14 to 7x7, generating a first down-sampled feature map of 256x7x7 dimensions. The down-sampling process preserves key information through feature compression, reduces redundant calculations, and provides higher semantic density feature input for the next level of attention calculation. The pre-set channel compression rate (512→256) is optimized through model training to minimize information loss.
[0166] The first down-sampled feature map is subjected to a window multi-head self-attention operation by a second window attention backbone block. The backbone block is pre-set with 8 attention heads, and the window size remains 7x7. Since the spatial resolution of the input feature map is reduced to 7x7, a single window now covers the entire feature map, essentially converting to a global self-attention calculation (but still retaining the window calculation form). Doubling the number of attention heads enhances the model's ability to analyze cross-regional features, capturing the distribution patterns of the insulating layer thickness in a larger range. This step outputs a second window feature map of 256x7x7 dimensions, whose features represent the insulating layer state information of a larger area on the cold plate surface.
[0167] The second window feature map is subjected to a down-sampling operation by a second down-sampling layer. The current 7x7 feature map is treated as a 2x2 window array (each window size is 3.5x3.5, rounded up to 4x4), and the features of adjacent 2x2 windows are concatenated along the channel dimension, increasing the number of channels to 1024 (256x4); then a 1x1 convolution kernel is used to compress the channels to 512, reducing the spatial resolution to 4x4, generating a second down-sampled feature map of 512x4x4 dimensions. This down-sampling further abstracts the global features, focusing on the core discriminative information of the overall uniformity of the insulating layer. The pre-set final resolution (4x4) is an empirical value that minimizes the computational load while preserving spatial information.
[0168] The second down-sampled feature map is subjected to a global self-attention operation by a third window attention backbone block. The backbone block is pre-set with 16 attention heads, and since the input feature map size is 4x4, a global self-attention mechanism (G-MSA) is directly used to calculate the fully connected dependency relationships between all 16 spatial positions. The multi-head design improves the model's ability to analyze complex features, such as simultaneously analyzing thickness mean, standard deviation, and local mutations, etc. This step outputs a third window feature map of 512x4x4 dimensions, whose features have integrated the global context information of the cold plate insulating layer, eliminating local noise interference.
[0169] A spatial dimension compression operation is performed on the third window feature map by a global average pooling layer. This operation calculates the average value along the spatial dimension for the 4x4 feature map of 512 channels respectively, compressing 16 feature values (4x4) of each channel to 1 scalar value, generating a 512-dimensional classification feature vector. This step eliminates the influence of spatial position difference on the classification result, and retains the semantic features of the channel dimension, such as different channels representing abstract properties such as insulation layer thickness, uniformity, interface integrity, etc.
[0170] A dimension mapping operation is performed on the 512-dimensional classification feature vector by a fully connected layer. The fully connected layer is preset with 1024 neurons, linearly transforms the input vector to a 1024-dimensional space through a weight matrix, and then introduces nonlinearity through a ReLU activation function. This operation maps high-dimensional features to a higher-dimensional hidden space, enhancing the model's fitting ability for complex classification boundaries. The preset hidden layer dimension (1024) is set according to the complexity of the classification task, ensuring sufficient accommodation of multi-dimensional discriminative information of the insulation layer state.
[0171] Based on the insulation layer classification vector, the insulation layer detection result of the target detection point is determined. The 1024-dimensional feature vector is input into the output layer (fully connected layer) and mapped to a dimension of the number of categories (such as 2 dimensions: normal or abnormal thickness, or 3 dimensions: uniform or locally non-uniform or overall non-uniform). The output value is converted to a probability distribution by the Softmax function, and the class corresponding to the maximum probability is selected as the detection result. This step completes the decision mapping from the feature space to the insulation layer state, and the output result contains quantitative information such as whether the insulation layer thickness of the detection point meets the standard and the uniformity level.
[0172] In some examples, based on the insulation layer detection result of the target detection point, the insulation performance detection result of the power battery cold plate is determined, including:
[0173] The number of detection points marked as "uniform" is counted among all the preset number of detection points (the number of detection points N represents N sensors, which is preset according to the size of the cold plate). The number of uniform detection points is divided by the total number of detection points, and then multiplied by 100% to obtain the uniformity proportion value of the uniform detection points. This value reflects the uniformity of the spatial distribution of the insulating layer on the surface of the cold plate. The thickness estimation values of all detection points (thickness detection results output by the insulating layer classification network) are extracted. The arithmetic mean of these thickness estimation values is calculated. For each thickness estimation value, the square of the deviation from the arithmetic mean is calculated. The square root of the average of all deviation squares is obtained to obtain the thickness standard deviation value. This value represents the range of discrete fluctuations of the thickness value of the insulating layer. The uniformity proportion threshold is preset according to the performance design requirements of the insulating paint (such as the thickness standard corresponding to the insulation resistance and voltage resistance level), which can be set to 85-90%, and the placement of the battery module area needs to be strictly controlled. It can also be adjusted flexibly according to the actual situation, and the specific limit is not limited in this patent. The thickness standard deviation threshold is preset according to the maximum fluctuation range of the insulating paint thickness, which can be set to ±10-50μm, and can be adjusted flexibly according to the actual situation, and the specific limit is not limited in this patent. Preferably, the uniformity proportion threshold of this patent is set to 85%, and the thickness standard deviation threshold is set to ±10μm.
[0174] When the average thickness of the cold plate insulating layer is ≥200μm, it is mapped to meet the basic insulating voltage resistance performance under DC 1000V discharge voltage (insulation resistance ≥500MΩ, voltage leakage current ≤1mA@DC 2700V, 60s); when the average thickness is ≥350μm, it is mapped to meet the enhanced insulating voltage resistance performance under DC 3800V discharge voltage (insulation resistance ≥500MΩ, voltage leakage current ≤1mA@DC 3800V, 60s). The average thickness is obtained by calculating the arithmetic mean of the thickness estimation values of all detection points. The uniform point proportion reflects the spatial coverage integrity of the insulating layer. When the proportion is ≥85% (preset uniformity proportion threshold), it indicates that the coating has no local defects and the mechanical strength is consistent, which can ensure the stability of the insulating performance under vibration, thermal cycle and other working conditions; if the proportion is <85%, there are coating shedding or thin area, which increases the risk of breakdown and easily causes electrical failure. The thickness standard deviation represents the dispersion degree of the thickness value. When the standard deviation is ≤±10μm (preset thickness standard deviation threshold), it indicates that the thickness fluctuation is within the allowable range, and the insulating safety margin is sufficient; if the standard deviation is >±10μm, the thickness range is too large, the thin area may cause electrical failure such as breakdown and leakage current, and the thick area may cause stress concentration and low heat transfer efficiency, and the insulating safety margin is insufficient.
[0175] When the cold plate is assembled in an 800V high-voltage battery pack, it needs to meet the enhanced insulation withstand voltage performance, the thickness average is ≥350μm, the uniform point proportion is ≥85% and the thickness standard deviation is ≤±10μm, at this time the cold plate can be safely applied to the 800V high-voltage battery system. When the cold plate is assembled in a 400V battery pack, it needs to meet the basic insulation withstand voltage performance, the thickness average is ≥200μm, the uniform point proportion is ≥85% and the thickness standard deviation is ≤±10μm, at this time the cold plate can be safely applied to the 400V high-voltage battery system. When the above two conditions are not met, process treatment or re-spraying or scrap treatment can be performed, which does not have the conditions for packaging, for example, when the thickness average is ≥200μm and the uniform point proportion is ≥85%, but the thickness standard deviation is >±10μm, although there is no coating damage in this state, the thickness fluctuation is out of standard, the detection point coordinates of the thickness out-of-standard (such as ±15μm) need to be marked, and local coating or increasing insulation pads or cold plate overall polishing is required after re-spraying, and ultrasonic detection is performed again until the thickness standard deviation is ≤±10μm, otherwise it is directly scrapped
[0176] It should be noted that the basic insulation withstand voltage performance refers to using >15N force to press the upper cover plate of the liquid cooling plate, using a withstand voltage tester, applying a DC 2700V discharge voltage, maintaining the voltage for 60s, no electric breakdown and spark phenomenon, and the withstand voltage leakage current is ≤1mA; applying a DC 1000V discharge voltage, maintaining the voltage for 60s, the insulation resistance is ≥500MΩ; the enhanced insulation withstand voltage performance refers to using >15N force to press the upper cover plate of the liquid cooling plate, using a withstand voltage tester, applying a DC 3800V discharge voltage, maintaining the voltage for 60s, no electric breakdown and spark phenomenon, and the withstand voltage leakage current is ≤1mA; applying a DC 1000V discharge voltage, maintaining the voltage for 60s, the insulation resistance is ≥500MΩ.
[0177] In some examples, the frequency domain feature vector, the time domain feature vector and the energy feature vector of the target detection point are subjected to feature fusion processing to generate a fusion feature vector of the target detection point, including:
[0178] The cosine similarity between each two feature vectors in the frequency domain feature vector, the time domain feature vector and the energy feature vector of the target detection point is calculated respectively;
[0179] Based on the comparison result of the preset similarity threshold and the cosine similarity, the target fusion feature set is determined;
[0180] The target fusion feature set is subjected to feature fusion processing to generate a fusion feature vector of the target detection point.
[0181] For example, the cosine similarity between each two types of feature vectors is calculated for the frequency domain feature vector, the time domain feature vector and the energy feature vector (dimension 1×196) of the target detection point. Specifically, it includes the similarity S freq_time, the similarity S of the time domain feature vector and the energy feature vector freq_energy , the similarity S of the time domain feature vector and the energy feature vector time_energy The cosine similarity is calculated by the ratio of the vector dot product and the modulus length product, and its value range is [-1, 1], which is used to quantify the direction consistency of different modal features: when the similarity tends to 1, it represents high redundancy of the features, tends to -1, it represents significant conflict, and tends to 0, it represents strong independence.
[0182] Based on the comparison results of the preset similarity threshold θ (which can be set to 0.7, and is optimized according to historical experimental data) and the three groups of cosine similarities, the target fusion feature set is dynamically determined. If there is at least one group of similarity greater than θ, the feature pair with the highest similarity is selected, the one with larger L2 norm is selected, and the third feature vector not participating in the pair is selected to form the target fusion feature set. If all similarities are less than or equal to θ, all three feature vectors are retained as the target fusion feature set. This step realizes the screening of redundant features through the preset threshold, reduces the computational complexity on the premise of ensuring the integrity of the information.
[0183] The determined target fusion feature set is fused. First, each feature vector is normalized to zero mean and unit variance to eliminate dimensional differences. Then, the normalized feature vectors are spliced along the channel dimension (2x196 dimensions when two features are retained, and 3x196 dimensions when three features are retained). Cross-channel weighted fusion is performed through a 1x1 convolution kernel to output a 1x196 dimensional convolution feature vector. After introducing non-linear transformation through a ReLU activation function, the importance weight of the feature position is calculated using a self-attention mechanism to generate a weighted feature vector. The weighted result and the original convolution feature vector are added by residual to avoid information degradation. Finally, a 1x196 dimensional fusion feature vector is generated through ReLU activation. This process ensures that the fusion features have both discriminability and robustness through multi-level operations such as normalization, splicing, convolution compression, attention weighting, and residual connection.
[0184] In summary, the embodiments of the present application quantify feature redundancy through cosine similarity, dynamically remove low-contribution features combined with a preset threshold, and reduce the computational complexity of the fusion process. For example, when the frequency domain and energy features are highly similar, only one of them is retained for fusion to avoid repeated calculations. At the same time, the cooperative design of the self-attention mechanism and the residual connection ensures that the fused feature vector accurately focuses on the key feature area of the insulation layer thickness anomaly. This strategy reduces the model inference time while ensuring detection accuracy, is particularly suitable for industrial scenarios with multiple detection points in parallel, provides an efficient solution for large-scale automated detection of power battery cold plates insulation layer, and suppresses noise interference through feature screening to improve the reliability of thickness and uniformity evaluation.
[0185] Please refer to Figure 2A structural schematic diagram of a detection device for insulation performance of a power battery cold plate is provided in an embodiment of the present application, and the device comprises:
[0186] A signal acquisition unit 21 is configured to acquire an ultrasonic echo signal of a target detection point on a surface of the power battery cold plate.
[0187] A graph generation unit 22 is configured to generate a frequency domain feature graph, a time domain feature graph and a time-frequency energy feature graph of the target detection point based on the ultrasonic echo signal of the target detection point.
[0188] A vector extraction unit 23 is configured to extract deep features corresponding to the frequency domain feature graph, the time domain feature graph and the time-frequency energy feature graph of the target detection point based on a graph feature extraction network, to obtain a frequency domain feature vector, a time domain feature vector and an energy feature vector of the target detection point.
[0189] A vector determination unit 24 is configured to perform feature fusion processing on the frequency domain feature vector, the time domain feature vector and the energy feature vector of the target detection point, to generate a fusion feature vector of the target detection point.
[0190] A local output unit 25 is configured to input the fusion feature vector of the target detection point into an insulation layer classification network, and output an insulation layer detection result of the target detection point.
[0191] An overall output unit 26 is configured to determine an insulation performance detection result of the power battery cold plate based on the insulation layer detection result of the target detection point.
[0192] Please refer to Figure 3 An electronic device 300 is also provided in an embodiment of the present application, which comprises a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and capable of running on the processor. The processor 320 implements the steps of the method for detecting the insulation performance of the power battery cold plate when executing the computer program 311.
[0193] Since the electronic device introduced in the embodiment is the device used to implement the detection device for the insulation performance of the power battery cold plate in the embodiment of the present application, the specific implementation mode of the electronic device in the embodiment and various changes thereof can be understood by those skilled in the art based on the method introduced in the embodiment of the present application. Therefore, how the electronic device implements the method in the embodiment of the present application is not described in detail here, and as long as the device used by those skilled in the art to implement the method in the embodiment of the present application belongs to the scope of the present application.
[0194] In the specific implementation process, the computer program 311 can implement any of the embodiments in the first aspect when executed by the processor.
[0195] It should be noted that in the above embodiments, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.
[0196] Those skilled in the art should understand that the embodiments of the present application can provide a method, a system or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-readable storage media containing computer-readable program code.
[0197] The embodiments of the present application are described with reference to flowcharts and / or block diagrams according to the methods, devices (systems) and computer program products of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams can be implemented by computer program instructions, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing devices to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for performing the functions specified in one or more blocks.
[0198] These computer program instructions can also be stored in a computer-readable memory capable of guiding a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including instruction means, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for performing the functions specified in one or more blocks.
[0199] These computer program instructions can also be loaded into a computer or other programmable data processing device, so that a series of operation steps are performed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for performing the functions specified in one or more blocks.
[0200] The embodiments of the present application also provide a computer program product, which includes computer software instructions, when the computer software instructions run on a processing device, so that the processing device executes Figure 1 The flow of the method for detecting the insulation performance of the power battery cold plate in the corresponding embodiment. The flow of the method for detecting the insulation performance of the power battery cold plate in the corresponding embodiment.
[0201] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on the computer, the processes or functions according to the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through wired or wireless mode. The computer-readable storage medium can be any available medium that can be stored by the computer or a data storage device such as a server, data center, etc. integrated with one or more available media. The available media can be magnetic media, optical media, or semiconductor media, etc.
[0202] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.
[0203] In several embodiments provided in the present application, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the apparatus embodiments described above are only schematic. The division of units is only a logical function division. Actual implementation can have another division manner. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, which can be electrical, mechanical or other form.
[0204] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or they can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.
[0205] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be in the form of hardware and / or software functional unit.
[0206] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for causing a computer device to perform all or part of the steps of the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory, a magnetic disk or an optical disk, and various program code storage media.
[0207] The above embodiments are only used to illustrate the technical solutions of the present application, but not limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.
[0208] Although the preferred embodiments of the present application have been described, those skilled in the art can make additional changes and modifications to the embodiments once they know the basic inventive concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.
[0209] Obviously, those skilled in the art can make various modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications of the present application fall within the scope of the claims of the present application and their equivalents, the present application also intends to include these modifications.
Claims
1. A method for detecting the insulation performance of a power battery cold plate, characterized in that, The method comprises: acquiring an ultrasonic echo signal of a target detection point on the surface of the power battery cold plate; generating a frequency domain feature map, a time domain feature map and a time-frequency energy feature map of the target detection point based on the ultrasonic echo signal of the target detection point; extracting deep features corresponding to the frequency domain feature map, the time domain feature map and the time-frequency energy feature map of the target detection point based on a feature map feature extraction network, to obtain a frequency domain feature vector, a time domain feature vector and an energy feature vector of the target detection point; performing feature fusion processing on the frequency domain feature vector, the time domain feature vector and the energy feature vector of the target detection point to generate a fusion feature vector of the target detection point; inputting the fusion feature vector of the target detection point into an insulation layer classification network to output an insulation layer detection result of the target detection point; determining an insulation performance detection result of the power battery cold plate based on the insulation layer detection result of the target detection point.
2. The method of claim 1, wherein, The method comprises: performing frequency domain transformation on the ultrasonic echo signal of the target detection point to generate an initial frequency spectrum of the target detection point; performing frequency spectrum refinement on a target frequency band based on the initial frequency spectrum of the target detection point and the target frequency band to generate an enhanced frequency spectrum of the target detection point; constructing a frequency domain feature matrix of the target detection point based on the enhanced frequency spectrum of the target detection point; performing graphical conversion on the frequency domain feature matrix of the target detection point to generate the frequency domain feature map of the target detection point.
3. The method of claim 1, wherein, The method comprises: constructing a polar coordinate mapping point set of the target detection point based on the ultrasonic echo signal of the target detection point; generating a symmetric point distribution map of the target detection point based on the polar coordinate mapping point set of the target detection point, and taking the symmetric point distribution map as the time domain feature map of the target detection point.
4. The method of claim 1, wherein, The method comprises: performing time-frequency transformation on the ultrasonic echo signal of the target detection point to generate a time-frequency spectrum energy matrix of the target detection point; extracting energy distribution features of the target detection point by performing frequency band decomposition on the time-frequency spectrum energy matrix of the target detection point through a preset filter bank; constructing a time-frequency energy distribution matrix of the target detection point based on the energy distribution features of the target detection point; converting the time-frequency energy distribution matrix of the target detection point into a visualized heat map, and taking the visualized heat map as the time-frequency energy feature map of the target detection point.
5. The method of claim 3, wherein, The method comprises: determining a signal maximum value, a signal minimum value and a current time point amplitude of the target detection point based on the ultrasonic echo signal of the target detection point; determining a lag time point amplitude of the target detection point based on a preset lag factor; performing normalization calculation on the current time point amplitude of the target detection point based on the signal maximum value and the signal minimum value to generate a normalized polar radius of the target detection point; The lag time point amplitude of the target detection point is normalized based on the signal maximum value and the signal minimum value, and a lag time point amplitude normalization quantity of the target detection point is generated; The lag time point amplitude of the target detection point is calculated based on the lag time point amplitude normalization quantity, a preset reference angle and a preset angle gain factor, and a first deflection angle and a second deflection angle of the target detection point are generated; The polar coordinate mapping point set of the target detection point is constructed based on the normalized polar radius, the first deflection angle and the second deflection angle.
6. The method of claim 1, wherein, The feature fusion processing of the frequency domain feature vector, the time domain feature vector and the energy feature vector of the target detection point generates a fusion feature vector of the target detection point, including: The frequency domain feature vector, the time domain feature vector and the energy feature vector of the target detection point are spliced along the channel dimension to generate a spliced feature vector of the target detection point; The spliced feature vector is convolved to generate a convolution feature vector of the target detection point; The convolution feature vector is weighted based on a self-attention mechanism to generate a weighted feature vector of the target detection point; The weighted feature vector and the convolution feature vector are residually fused to generate the fusion feature vector of the target detection point.
7. The method of claim 1, wherein, The feature fusion processing of the frequency domain feature vector, the time domain feature vector and the energy feature vector of the target detection point generates a fusion feature vector of the target detection point, including: The cosine similarity between each two feature vectors of the frequency domain feature vector, the time domain feature vector and the energy feature vector of the target detection point is calculated respectively; The target fusion feature set is determined based on the comparison result of the preset similarity threshold and the cosine similarity; The target fusion feature set is subjected to feature fusion processing to generate the fusion feature vector of the target detection point.
8. A device for detecting the insulation performance of a power battery cold plate, characterized in that, including: A signal acquisition unit is configured to acquire an ultrasonic echo signal of a target detection point on the surface of a power battery cold plate; A graph generation unit is configured to generate a frequency domain feature graph, a time domain feature graph and a time-frequency energy feature graph of the target detection point based on the ultrasonic echo signal of the target detection point; A vector extraction unit is configured to extract deep features corresponding to the frequency domain feature graph, the time domain feature graph and the time-frequency energy feature graph of the target detection point based on a graph feature extraction network, and obtain a frequency domain feature vector, a time domain feature vector and an energy feature vector of the target detection point; A vector determination unit is configured to perform feature fusion processing on the frequency domain feature vector, the time domain feature vector and the energy feature vector of the target detection point, and generate a fusion feature vector of the target detection point; A local output unit is configured to input the fusion feature vector of the target detection point into an insulation layer classification network, and output an insulation layer detection result of the target detection point; An overall output unit is configured to determine an insulation performance detection result of the power battery cold plate based on the insulation layer detection result of the target detection point.
9. An electronic device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor is configured to implement the steps of the method for detecting the insulation performance of a power battery cold plate according to any one of claims 1 to 7 when executing the computer program stored in the memory.
10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is configured to implement the method for detecting the insulation performance of a power battery cold plate according to any one of claims 1 to 7 when executed by the processor.
Citation Information
Cited By
System and method for detecting the state of a two-phase cold plate liquid cooling system
CN122430091A
System and method for detecting the state of a two-phase cold plate liquid cooling system
CN122430091B