import numpy as np

import tensorflow as tf

from tensorflow.keras import layers, models

def build_crash_detector(input_shape):

"""

Builds a Bi-Directional LSTM for detecting high-impact anomalies.

Input Shape: (window_size, features) - e.g., (50, 6) for 50Hz data

Features: acc_x, acc_y, acc_z, gyro_x, gyro_y, gyro_z

"""

model = models.Sequential([

# Bi-directional layer to capture pre-impact and impact context

layers.Bidirectional(layers.LSTM(64, return_sequences=True), input_shape=input_shape),

layers.Dropout(0.2),

layers.Bidirectional(layers.LSTM(32)),

layers.Dense(32, activation='relu'),

# Output: Probability of a crash (0 to 1)

layers.Dense(1, activation='sigmoid')

])

model.compile(optimizer='adam',

loss='binary_crossentropy',

metrics=['accuracy', tf.keras.metrics.Precision(), tf.keras.metrics.Recall()])

return model

# Example dimensions: 1 second of data at 50Hz with 6 sensor axes

input_dim = (50, 6)

crash_ai = build_crash_detector(input_dim)

crash_ai.summary()

import tensorflow as tf

from tensorflow.keras import layers, Model

def build_advanced_crash_ai(imu_shape=(100, 6), gps_shape=(10, 2)):

"""

Advanced Multimodal Fusion Model

- IMU: 100 samples (2 seconds @ 50Hz) of Accel(3) + Gyro(3)

- GPS: 10 samples of Speed + Heading

"""

# 1. IMU Branch (Temporal Analysis)

imu_input = layers.Input(shape=imu_shape, name="imu_data")

x = layers.Bidirectional(layers.LSTM(64, return_sequences=True))(imu_input)

# Self-Attention Layer to pinpoint the impact moment

query = layers.Dense(128)(x)

value = layers.Dense(128)(x)

attention = layers.Attention()([query, value])

imu_feat = layers.GlobalAveragePooling1D()(attention)

# 2. GPS Branch (Contextual Analysis)

gps_input = layers.Input(shape=gps_shape, name="gps_data")

gps_feat = layers.Flatten()(gps_input)

gps_feat = layers.Dense(16, activation='relu')(gps_feat)

# 3. Fusion Layer

combined = layers.Concatenate()([imu_feat, gps_feat])

z = layers.Dense(64, activation='swish')(combined) # Swish is smoother for gradients

z = layers.Dropout(0.3)(z)

# Output: 0 (Normal), 1 (Minor Impact), 2 (Severe Crash)

output = layers.Dense(3, activation='softmax', name="crash_severity")(z)

model = Model(inputs=[imu_input, gps_input], outputs=output)

model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])

return model

crash_ai = build_advanced_crash_ai()