A radio frequency system testability modeling method based on improved auto-encoder

Abstract

This invention discloses a testability modeling method for radio frequency (RF) systems based on an improved autoencoder, belonging to the field of RF system testability modeling. The method involves establishing a simulation circuit model, acquiring and expanding measurement data, using a voltage drop threshold to perform fuzzy group annotation on the expanded measurement data to obtain a multi-valued D matrix, establishing an autoencoder model, and initializing its parameters, training the simulation circuit model, and finally outputting a reconstructed matrix. This invention incorporates a fuzzy group constraint loss function into the autoencoder, resulting in an improved autoencoder. This improved autoencoder can learn the intrinsic structure of the data, discover data patterns through training data, and correlate fault fuzzy groups with latent representations when processing analog circuit test signal data, thereby verifying the rationality of the fuzzy group partitioning.

Description

Technical Field

[0001] This invention belongs to the field of radio frequency system testability modeling, and is a testability modeling method for radio frequency systems based on an improved autoencoder. Background Technology

[0002] Integrated radio frequency (RF) systems, supported by RF integration, software integration, and aperture integration technologies, aim to solve the problem of space constraints in equipment systems. They are shared electronic platforms that highly integrate radar, electronic warfare, and communication functions. Currently, they are widely used in military electronic countermeasures, target identification, and navigation, as well as civilian search and rescue, air traffic control, and satellite communications. To ensure the stable operation of integrated RF systems in complex environments, testing, diagnosis, and support must be comprehensively considered from the initial design and development stages. This requires the equipment to possess good in-flight testing and self-diagnostic capabilities, and to provide a good and convenient interface for external testing—that is, testability. Currently, the testability design for complex systems still faces a difficult technical process and key technical problem: poor accuracy in fault detection and diagnosis, long testing and diagnosis times, low efficiency, high costs, and wasted maintenance and support resources. These issues seriously hinder the rapid formation of combat and support capabilities and increase the total life-cycle cost of the equipment.

[0003] As the core of testability design, testability modeling and analysis can objectively evaluate the testability design level of a system, which is of great significance for optimizing testability design schemes, reducing testing time and costs, and improving testing efficiency. Many research institutes and universities in my country started relatively late in the field of testability research, and have carried out a great deal of work in theoretical research and application development. Although they have achieved certain results, there is still a gap between China and foreign countries in the modeling technology of multi-layered complex systems, especially in the research on testability modeling methods for complex airborne systems, where the amount of related research is still relatively small. This is mainly reflected in the modeling and analysis of multi-valued attribute systems. Multi-valued attribute systems have unique properties; the elements within the system are often not simple binary logic but have multiple possible values, which requires consideration of more complex factors during the modeling process.

[0004] For the radio frequency (RF) circuits targeted by this patent, a multi-valued attribute system based on fuzzy group integer coding is often used, which represents the values ​​of matrix elements as fuzzy group numbers for different faults. For RF circuits, a variance-based analog circuit fault isolation algorithm is generally employed, and this variance is typically set to -0.7V of the diode's voltage drop under normal bias. However, using only the 0.7V voltage drop as the boundary for fuzzy groups has significant limitations. In actual analog circuit systems, voltage signal changes are influenced by numerous interacting factors, including the parameter dispersion of different components, differences in temperature and humidity in the circuit's operating environment, and variations in electromagnetic interference intensity. This single threshold setting cannot fully capture the changing patterns of voltage data under complex operating conditions, potentially leading to many misjudgments and affecting the accurate assessment of the circuit's state. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides a testability modeling method for radio frequency (RF) systems based on an improved autoencoder, belonging to the field of RF system testability modeling. The method involves establishing a simulation circuit model, acquiring measurement data, and expanding the measurement data; using a voltage drop threshold to perform fuzzy group annotation on the expanded measurement data to obtain a multi-valued D matrix; establishing an autoencoder model and initializing its parameters; training the simulation circuit model; and finally outputting the reconstructed matrix. This invention incorporates a fuzzy group constraint loss function into the autoencoder, resulting in an improved autoencoder. This improved autoencoder can learn the intrinsic structure of the data, discover data patterns through learning from training data, and associate fault fuzzy groups with latent representations when processing analog circuit test signal data, thereby verifying the rationality of the fuzzy group partitioning.

[0006] A testability modeling method for RF systems based on an improved self-encoder, such as Figure 1 As shown, it includes the following steps:

[0007] Step 1: Establish a simulation circuit model using the Simulation Program with Integrated Circuit Emphasis (SPICE) protocol to obtain voltage data at multiple test points; expand the voltage data by setting tolerances to cause resistance value fluctuations.

[0008] Step 2: Construct an adaptive fuzzy group-constrained autoencoder model;

[0009] Step 3: Using the threshold prediction branch of the adaptive fuzzy group constrained autoencoder model, the optimal fuzzy group threshold is self-learned for each test point, and the expanded voltage data is fuzzy group labeled to obtain the original fuzzy group label.

[0010] Step 4: Initialize the adaptive fuzzy group-constrained autoencoder model parameters, including weights, biases, and balance parameters λ.

[0011] Step 5: Train an adaptive fuzzy group-constrained autoencoder model using the augmented voltage data; during the training process, dynamically adjust the balance parameter λ based on the ratio of reconstruction quality loss to adaptive fuzzy group inconsistency loss.

[0012] Step 6: Output the reconstruction matrix using the trained adaptive fuzzy group constrained autoencoder model; by rounding down the average of multiple fuzzy group integer codes for each test point, the integrated and simplified reconstruction matrix is ​​obtained, which is the final multi-valued D matrix.

[0013] Furthermore, in step two, the adaptive fuzzy group-constrained autoencoder model includes an encoder, a decoder, and an adaptive loss function;

[0014] The encoder includes a feature extraction subnetwork and a test point threshold prediction subnetwork; the feature extraction subnetwork sequentially includes a first hidden layer, a second hidden layer, and a latent representation layer.

[0015] The decoder comprises, in sequence, a first decoding hidden layer, a second decoding hidden layer, and a reconstructed output layer;

[0016] The adaptive loss function includes reconstruction quality loss and adaptive fuzzy group inconsistency loss.

[0017] Furthermore, in step four, the parameters of the fuzzy group-constrained autoencoder model are initialized as follows:

[0018] For the weights and biases of each layer of the encoder, a uniform distribution initialization method is used; for layers containing n... in input neurons and n out The weight matrix W is then applied to the l-th layer of the output neurons. l element W la a = 1, ..., n in ;

[0019] The weight matrix W l Initialize to:

[0020]

[0021] The bias vector b l Initialize to:

[0022]

[0023] Where U() represents a uniform distribution.

[0024] Furthermore, in step five, the training process of the adaptive fuzzy group-constrained autoencoder model is as follows:

[0025] Input dataset of adaptive fuzzy group-constrained autoencoder model n is the number of training samples, and M is the number of test points; the original voltage data in dataset X. This represents the M effective voltage values ​​of the i-th sample. Let x represent the set of real numbers. i =[x i1 ,x i2 ,...x ij ...,x iM ], where x ij Let j be the j-th valid value of the i-th sample, where j = 1, 2, ..., M;

[0026] Step 5.1: Extract the raw voltage data x using the encoder. i Low-dimensional features;

[0027] Step 5.1.1, convert the original voltage data x i The input is fed into the first hidden layer of the feature extraction sub-network, and the output is the preliminary feature. The i-th sample is in the j-th hidden layer. 1 The output of each neuron for:

[0028]

[0029] Where, j 1 =1,2,...,64; Let be the weight matrix of the first hidden layer, and let j be the weight matrix of the first hidden layer. 1 The element in row k and column j corresponds to the kth test point. 1 The connection weights of each neuron; x ik It is the original voltage data x i The kth element; The j-th bias vector of the first hidden layer 1 1 element; sigmoid activation function σ sigmoid ()for:

[0030]

[0031] Where z is the independent variable of the sigmoid activation function; e is the natural constant;

[0032] Step 5.1.2, h 1i The input is fed into the second hidden layer of the feature extraction subnetwork, and the output is the deep feature layer. The i-th sample is in the second hidden layer at the j-th position. 2 The output of each neuron for:

[0033]

[0034] in, It is the kth output of the first hidden layer 1 One element, This is the weight matrix of the second hidden layer. This is the bias vector for the second hidden layer;

[0035] Step 5.1.3, h 2i The input is fed into the latent representation layer of the feature extraction subnetwork, and the output is a low-dimensional latent representation. z i Used for subsequent decoder reconstruction and threshold prediction, the i-th sample is in the j-th position of the latent representation layer. 3 The output of each neuron for:

[0036]

[0037] in, It is the kth output of the second hidden layer 2 One element, Here is the weight matrix of the latent representation layer. The bias vector of the potential representation layer;

[0038] Step 5.1.4, convert the low-dimensional latent representation z i The input is fed into the threshold prediction subnetwork at the test point, and the output is the fuzzy group threshold τ corresponding to the test point of the sample. ik' ;

[0039] The fuzzy group threshold τ corresponding to the M test points of the i-th sample i =[τ i1 ,τ i2 ,...,τ iM ], τ ik' τ is the optimal fuzzy group threshold for the i-th sample and the k'-th test point. ik' for:

[0040]

[0041] in, It is the k-th latent representation of the low-dimensional dimension 3 One element, The weight matrix for the threshold prediction branch. σ is the bias vector for the threshold prediction branch. ReLU (z) = max(0,z) is the activation function, ensuring that the threshold is non-negative;

[0042] Step 5.2, the decoder will process the raw voltage data x i From the low-dimensional latent representation z i Reconstructing voltage data Ensure voltage data is reconstructed Compared with the original voltage data x i The voltage characteristics are consistent, resulting in a multi-valued D matrix;

[0043] Step 5.2.1, convert the low-dimensional latent representation z i Input to the first decoding hidden layer, output The i-th sample is in the j-th hidden layer of the first decoding layer. 4 The output of each neuron

[0044]

[0045] in, It is the k-th latent representation of the low-dimensional dimension 4 One element, This is the weight matrix of the first decoding hidden layer. This is the bias vector of the first decoding hidden layer; Let x be the activation function, and let x be the independent variable of the activation function.

[0046] Step 5.2.2, will The input is fed into the second decoding hidden layer, and the output is... The i-th sample is in the j-th hidden layer of the second decoding layer. 5 The output of each neuron for:

[0047]

[0048] in, It is the kth output of the first decoding hidden layer 5 One element, This is the weight matrix of the second decoding hidden layer. This is the bias vector for the second decoding hidden layer;

[0049] Step 5.2.3, will Input to the reconstruction output layer, output reconstruction voltage data The i-th sample is in the j-th reconstructed output layer. 6 The output of each neuron for:

[0050]

[0051] in, To reconstruct the weight matrix of the output layer; To reconstruct the bias vector of the output layer; σ Linear(z) = z is a linear activation function that preserves the original amplitude information of the voltage data;

[0052] Step 5.3, by reconstructing the mass loss L recon And adaptive fuzzy group inconsistency loss L fuzzy Calculate the adaptive loss function L;

[0053] Step 5.3.1, calculate the reconstruction mass loss L recon :

[0054] Reconstruction mass loss L recon Measuring raw voltage data x i With reconstructed voltage data The difference is expressed using the mean squared error (MSE), as follows:

[0055]

[0056] Where nM is the total amount of data, ensuring that the loss value is independent of the number of samples and the number of test points;

[0057] Step 5.3.2, calculate the adaptive fuzzy group inconsistency loss L. fuzzy :

[0058] Adaptive fuzzy group inconsistency loss L fuzzy Self-learning threshold τ based on each test point ij Calculate the inconsistency ratio between the original fuzzy group labels and the reconstructed fuzzy group labels:

[0059] The method for generating the original fuzzy group labels is as follows:

[0060] For the original voltage data x i According to the self-learning threshold τ ik' Divide into fuzzy groups;

[0061] If x ij ∈[V min,j +kτ ik' V min,j +(k+1)τ ik' ], k = 0, 1, ..., V min,j V min,j If the minimum voltage at the j-th effective value point is y, then the original fuzzy group label y ij =k;

[0062] The method for generating reconstructed fuzzy group labels is as follows:

[0063] For reconstructed voltage data Generate reconstruction labels

[0064] Adaptive fuzzy group inconsistency loss Lfuzzy :

[0065]

[0066] Where I() represents an indicator function;

[0067] Step 5.3.3, calculate the adaptive loss function L:

[0068] L = L recon +λL fuzzy ;

[0069] Wherein, the equilibrium parameter λ is based on L recon and L fuzzy The ratio is adjusted; the adaptive fuzzy group constraint autoencoder model is continuously trained by dynamically adjusting the balance parameter λ.

[0070]

[0071] Where, λ base The static baseline value is optimized based on experiments, and the range of λ is limited to [0.3, 2.0] to avoid extreme values ​​causing model collapse.

[0072] Furthermore, in step six, the process for obtaining the final multi-valued D matrix is ​​as follows:

[0073] Step 6.1, forward propagation and calculation of the adaptive loss function L:

[0074] Repeat steps 1 through 5 to obtain the low-dimensional latent representation z. i fuzzy group threshold τ i and reconstruct voltage data and reconstruction mass loss L recon Adaptive fuzzy group inconsistency loss L fuzzy 1. Balancing parameter λ and adaptive loss function L;

[0075] Step 6.2, Backpropagation and Parameter Update:

[0076] The gradient of the adaptive loss function L with respect to the model parameters is calculated using the chain rule, and the parameters are updated using gradient descent. The formula for updating the weights is as follows:

[0077]

[0078] In the formula, α is the learning rate; It is an adaptive loss function L with respect to weights W i gradient; W old This is the weight matrix before the update; W new This is the updated weight matrix;

[0079] For the bias vector bi Update formula:

[0080]

[0081] In the formula, b old It is the bias vector before the update; b new It is the updated bias vector;

[0082] Step 6.3, repeat forward and backward propagation; when L fluctuates less than ω after 10 consecutive training rounds, ω = 1e-5, stop training; or stop training when the maximum number of training rounds is reached, where the maximum number of training rounds is in the range of [100, 500].

[0083] The technical effects of this invention are as follows:

[0084] This application adds an adaptive fuzzy group consistency constraint to the reconstruction quality constraint, enabling the adaptive fuzzy group constraint autoencoder model to prioritize optimization of weaknesses at different training stages. The fuzzy group integer encoding matrix proposed in this invention is more reasonable than manual partitioning. Compared with conventional machine learning algorithms, this invention solves the problems of poor robustness and insufficient adaptability caused by traditional methods relying on manually preset thresholds and static parameters by allowing the model to learn the optimal fuzzy group threshold for each test point and dynamically adjusting the balance parameters based on the loss ratio. This improves the accuracy and engineering adaptability of multi-value testability modeling of RF systems. Attached Figure Description

[0085] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0086] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0087] This invention uses an active filter circuit as the experimental object. Active filters are an important component of radio frequency (RF) circuits, which can improve the signal-to-noise ratio and stability of the system through gain compensation, precise frequency selection, and impedance matching. Moreover, active filter circuits are a type of RF circuit. Therefore, using active filter circuits as the experimental object has important reference value for the design of testability modeling methods for RF systems.

[0088] Using an active power filter circuit as the experimental object, the effective voltage values ​​under 19 operating conditions were measured at 11 test points. This embodiment selects 19 fault types at 11 test points: T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, and T11; and the fault types are F0, F1, F2, F3, F4, F5, F6, F7, F8, F9, F10, F11, F12, F13, F14, F15, F16, F17, F18, and F19. Based on the relative magnitudes of the resistances in the circuit, low resistance values ​​(e.g., 1Ω) represent short circuits, and high resistance values ​​(e.g., 10MΩ) represent open circuits. A tolerance of ±0.2% is set to allow for slight fluctuations in the resistance values ​​of the series open / short circuit resistors, thus expanding the data. The final effective voltage values ​​of the active power filter circuit are shown in Table 1.

[0089] Table 1. Effective voltage values ​​of the active filter circuit

[0090]

[0091]

[0092] The fuzzy group integer encoding matrix of the active filter circuit obtained after step six is ​​shown in Table 2:

[0093] Table 2. Fuzzy group integer coding matrix of active filter circuit.

[0094]

[0095] After integration and simplification, the final multi-valued D matrix is ​​shown in Table 3:

[0096] Table 3 Final Multivalued D Matrix

[0097]

[0098]

[0099] As shown in Table 3, T2 and T3 in the fuzzy group integer coding moments based on the improved autoencoder are still redundant tests. One column needs to be deleted when constructing the final multi-valued D matrix, which is consistent with the result obtained from the fuzzy group coding matrix in step two. The fault isolation rate reaches 100%, and this model does not require manual adjustment of the threshold and λ, adapting to the voltage distribution differences at different test points and under different operating conditions in the RF system, thus reducing parameter debugging costs in engineering applications. The above results fully demonstrate the effectiveness and superiority of fuzzy group coding based on the improved autoencoder.

Claims

1. A testability modeling method for radio frequency systems based on an improved self-encoder, characterized in that, The radio frequency system testability modeling method includes the following steps: Step 1: Establish a simulation circuit model based on the integrated circuit simulation program protocol and obtain voltage data at multiple test points; expand the voltage data by setting tolerances to cause resistance value fluctuations. Step 2: Construct an adaptive fuzzy group-constrained autoencoder model; Step 3: Using the threshold prediction branch of the adaptive fuzzy group constrained autoencoder model, the optimal fuzzy group threshold is self-learned for each test point, and the expanded voltage data is fuzzy group labeled to obtain the original fuzzy group label. Step 4: Initialize the adaptive fuzzy group-constrained autoencoder model parameters, including weights, biases, and balance parameters λ. Step 5: Train an adaptive fuzzy group-constrained autoencoder model using the augmented voltage data; during the training process, dynamically adjust the balance parameter λ based on the ratio of reconstruction quality loss to adaptive fuzzy group inconsistency loss. Step 6: Output the reconstruction matrix using the trained adaptive fuzzy group constrained autoencoder model; by rounding down the average of multiple fuzzy group integer codes for each test point, the integrated and simplified reconstruction matrix is ​​obtained, which is the final multi-valued D matrix.

2. The testability modeling method for an RF system based on an improved autoencoder according to claim 1, characterized in that, In step two, the adaptive fuzzy group-constrained autoencoder model includes an encoder, a decoder, and an adaptive loss function; The encoder includes a feature extraction subnetwork and a test point threshold prediction subnetwork; The feature extraction subnetwork includes, in sequence, a first hidden layer, a second hidden layer, and a latent representation layer; The decoder comprises, in sequence, a first decoding hidden layer, a second decoding hidden layer, and a reconstructed output layer; The adaptive loss function includes reconstruction quality loss and adaptive fuzzy group inconsistency loss.

3. The testability modeling method for an RF system based on an improved autoencoder according to claim 1, characterized in that, In step four, the parameters of the fuzzy group-constrained autoencoder model are initialized as follows: For the weights and biases of each layer of the encoder, a uniform distribution initialization method is used; for layers containing n... in input neurons and n out The weight matrix W is then applied to the l-th layer of the output neurons. l element W la a = 1, ..., n in ; The weight matrix W l Initialize to: The bias vector b l Initialize to: Where U() represents a uniform distribution.

4. The testability modeling method for an RF system based on an improved autoencoder according to claim 1, characterized in that, In step five, the training process of the adaptive fuzzy group-constrained autoencoder model is as follows: Input dataset of adaptive fuzzy group-constrained autoencoder model n is the number of training samples, and M is the number of test points; the original voltage data in dataset X. This represents the M effective voltage values ​​of the i-th sample. Let x represent the set of real numbers. i =[x i1 ,x i2 ,...x ij ...,x iM ], where x ij Let j be the j-th valid value of the i-th sample, where j = 1, 2, ..., M; Step 5.1: Extract the raw voltage data x using the encoder. i Low-dimensional features; Step 5.1.1, convert the original voltage data x i The input is fed into the first hidden layer of the feature extraction sub-network, and the output is the preliminary feature. The i-th sample is in the j-th hidden layer. 1 The output of each neuron for: Where, j 1 =1,2,...,64; Let be the weight matrix of the first hidden layer, and let j be the weight matrix of the first hidden layer. 1 The element in row k and column j corresponds to the kth test point. 1 The connection weights of each neuron; x ik It is the original voltage data x i The kth element; The j-th bias vector of the first hidden layer 1 1 element; sigmoid activation function σ sigmoid ()for: Where z is the independent variable of the sigmoid activation function; e is the natural constant; Step 5.1.2, h 1i The input is fed into the second hidden layer of the feature extraction subnetwork, and the output is the deep feature layer. The i-th sample is in the second hidden layer at the j-th position. 2 The output of each neuron for: in, It is the kth output of the first hidden layer 1 One element, This is the weight matrix of the second hidden layer. This is the bias vector for the second hidden layer; Step 5.1.3, h 2i The input is fed into the latent representation layer of the feature extraction subnetwork, and the output is a low-dimensional latent representation. z i Used for subsequent decoder reconstruction and threshold prediction, the i-th sample is in the j-th position of the latent representation layer. 3 The output of each neuron for: in, It is the kth output of the second hidden layer 2 One element, Here is the weight matrix of the latent representation layer. The bias vector of the potential representation layer; Step 5.1.4, convert the low-dimensional latent representation z i The input is fed into the threshold prediction subnetwork at the test point, and the output is the fuzzy group threshold τ corresponding to the test point of the sample. ik' ; The fuzzy group threshold τ corresponding to the M test points of the i-th sample i =[τ i1 ,τ i2 ,...,τ iM ], τ ik' τ is the optimal fuzzy group threshold for the i-th sample and the k'-th test point. ik' for: in, It is the k-th latent representation of the low-dimensional dimension 3 One element, The weight matrix for the threshold prediction branch. σ is the bias vector for the threshold prediction branch. ReLU (z) = max(0,z) is the activation function, ensuring that the threshold is non-negative; Step 5.2, the decoder will process the raw voltage data x i From the low-dimensional latent representation z i Reconstructing voltage data Ensure voltage data is reconstructed Compared with the original voltage data x i The voltage characteristics are consistent, resulting in a multi-valued D matrix; Step 5.2.1, convert the low-dimensional latent representation z i Input to the first decoding hidden layer, output The i-th sample is in the j-th hidden layer of the first decoding layer. 4 The output of each neuron Among them, z ik 4 It is the k-th latent representation of the low-dimensional dimension 4 One element, This is the weight matrix of the first decoding hidden layer. This is the bias vector of the first decoding hidden layer; Let x be the activation function, and let x be the independent variable of the activation function. Step 5.2.2, will The input is fed into the second decoding hidden layer, and the output is... The i-th sample is in the j-th hidden layer of the second decoding layer. 5 The output of each neuron for: in, It is the kth output of the first decoding hidden layer 5 One element, This is the weight matrix of the second decoding hidden layer. This is the bias vector for the second decoding hidden layer; Step 5.2.3, will Input to the reconstruction output layer, output reconstruction voltage data The i-th sample is in the j-th reconstructed output layer. 6 The output of each neuron for: in, To reconstruct the weight matrix of the output layer; To reconstruct the bias vector of the output layer; σ Linear (z) = z is a linear activation function that preserves the original amplitude information of the voltage data; Step 5.3, by reconstructing the mass loss L recon And adaptive fuzzy group inconsistency loss L fuzzy Calculate the adaptive loss function L; Step 5.3.1, calculate the reconstruction mass loss L recon : Reconstruction mass loss L recon Measuring raw voltage data x i With reconstructed voltage data The difference is expressed using the mean squared error (MSE), as follows: Step 5.3.2, calculate the adaptive fuzzy group inconsistency loss L. fuzzy : Adaptive fuzzy group inconsistency loss L fuzzy Self-learning threshold τ based on each test point ij Calculate the inconsistency ratio between the original fuzzy group labels and the reconstructed fuzzy group labels: The method for generating the original fuzzy group labels is as follows: For the original voltage data x i According to the self-learning threshold τ ik' Divide into fuzzy groups; If x ij ∈[V min,j +kτ ik' V min,j +(k+1)τ ik' ], k = 0, 1, ..., V min,j V min,j If the minimum voltage at the j-th effective value point is y, then the original fuzzy group label y ij =k; The method for generating reconstructed fuzzy group labels is as follows: For reconstructed voltage data Generate reconstruction labels Adaptive fuzzy group inconsistency loss L fuzzy : Where I() represents an indicator function; Step 5.3.3, calculate the adaptive loss function L: L=L recon +λL fuzzy ; Wherein, the equilibrium parameter λ is based on L recon and L fuzzy The ratio is adjusted; the adaptive fuzzy group constraint autoencoder model is continuously trained by dynamically adjusting the balance parameter λ. Where, λ base The static baseline value is optimized based on experiments, and the range of λ is limited to [0.3, 2.0] to avoid extreme values ​​causing model collapse.

5. The testability modeling method for an RF system based on an improved autoencoder according to claim 1, characterized in that, In step six, the process of obtaining the multi-valued D matrix is ​​as follows: Step 6.1, forward propagation and calculation of the adaptive loss function L: Repeat steps 1 through 5 to obtain the low-dimensional latent representation z. i fuzzy group threshold τ i and reconstruct voltage data and reconstruction mass loss L recon Adaptive fuzzy group inconsistency loss L fuzzy 1. Balancing parameter λ and adaptive loss function L; Step 6.2, Backpropagation and Parameter Update: The gradient of the adaptive loss function L with respect to the model parameters is calculated using the chain rule, and the parameters are updated using gradient descent. The formula for updating the weights is as follows: In the formula, α is the learning rate; It is an adaptive loss function L with respect to weights W i gradient; W old This is the weight matrix before the update; W new This is the updated weight matrix; For the bias vector b i Update formula: In the formula, b old It is the bias vector before the update; b new It is the updated bias vector; Step 6.3, repeat forward and backward propagation; when L fluctuates less than ω after 10 consecutive training rounds, ω = 1e-5, stop training; or stop training when the maximum number of training rounds is reached, where the maximum number of training rounds is in the range of [100, 500].

6. An electronic device, characterized in that, include: One or more processors; Memory; One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs being configured to perform the method as described in any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program code that can be invoked by a processor to execute the method as described in any one of claims 1 to 5.

Images